You are on page 1of 237

Improved Turbulence Models for

Computational Wind Engineering


by Gary Easom B.Eng., M.Res.
Thesis submitted to the University of Nottingham for the degree of Doctor of
Philosophy, January 2000.
Supervisor
Dr N.G. Wright School of Civil Engineering, The University of Nottingham.
Associate Supervisor
Dr R.P. Hoxey Bio-engineering Division, Silsoe Research Institute.
ii
Table of contents.
- Abstract. vii
- List of figures viii
- List of tables. .... xv
- Acknowledgements. xvi
- List of Publications. ........................................................................ xvii
1. Introduction ........................................................................ 1
2. Wind Engineering
2.1 - Introduction. 5
2.2 - Wind Engineering Bluff body Aerodynamics. 5
2.2.1 - Fundamental Aspects. 6
2.2.2 - Bluff Body Flow Fields. 8
2.2.3 - Codes of Practice. 10
2.2.4 - Wind Damage Costs. 11
2.3 - Computational Wind Engineering. 11
2.3.1 - The Beginnings of CWE 1980s and early 1990s. 12
3. Solution Procedures and Numerical Analysis
3.1 - Introduction. 15
3.2 - The Navier-Stokes Equations. 15
3.3 - The Reynolds Stresses. 17
3.4 - Common Discretisation Schemes. 19
3.5 - Common Differencing Schemes Discretisation
of the Convection Term. 22
3.5.1 - The Central Differencing Scheme. 24
3.5.2 - The Upwind Differencing Scheme. 24
3.5.3 - The Hybrid Differencing Scheme. 25
3.5.4 - The QUICK Differencing Scheme. 25
3.5.5 - The CCCT Differencing Scheme. 26
3.6 - Calculation of the Flow Field and Pressure Equation. . 27
3.6.1 - Solving the Simultaneous Equations. 27
3.7 - Summary. 28
iii
4. Turbulence Modelling
4.1 - Introduction. 29
4.2 - Reynolds Averaged Navier-Stokes
Equations. 30
4.2.1 - The Eddy Viscosity Concept. 30
4.2.2 - The Mixing Length Model. 32
4.2.3 - The Standard k - Turbulence Model. 32
4.2.3.1 - Discussion. 34
4.2.4 - The Low Reynolds number k - Turbulence Model. . 35
4.2.5 - The k - Equation Model. 36
4.2.6 - The Renormalisation Group (RNG)
k- Turbulence Model. 37
4.2.6.1 - Discussion. 38
4.2.7 - The Differential Stress Equation Model (DSM). 39
4.2.7.1 - Discussion. 41
4.2.8 - The Algebraic Stress Model. 43
4.2.8.1 - Discussion. 43
4.3 - Turbulent Wall Boundary Conditions. 43
4.3.1 - The Linear Sub-Layer. 45
4.3.2 - The Log Law Layer. 46
4.4 - Alternatives to the Navier-Stokes Equations. 46
4.4.1 - Chaos. 46
4.4.2 - The Discrete Vortex Method. 47
4.5 Summary. 47
5. Developments in Turbulence Modelling
5.1 - Introduction. 48
5.2 - The MMK k - Turbulence Model. 48
5.3 - Two Layer Turbulence Models. 51
5.3.1 - Discussion. 52
5.4 - Revisions to the Boussinesq Hypothesis -
The Non-Linear k- Model. 53
5.5 - Wall Reflection Terms / The Pressure Strain
Relationship for the Differential Stress Model. 58
5.5.1 - Theory. 58
5.5.2 - Modelling. 58
5.6 - Direct Numerical Simulation. 61
5.7 - Large Eddy Simulation. 62
iv
5.7.1 - Discussion. .. 63
5.7.2 - Coherent Structure Capturing or Very Large Eddy
Simulation. .. 64
5.7.3 - Discussion. .. 64
5.8 - Summary. .. 65
6. Results: A Comparison of Data Obtained from Both CFD Simulations
and Full Scale Experimental Studies.
6.1 - Introduction. .. 66
6.2 - Full-Scale Test Data. .. 67
6.2.1 - The 6m Cube. .. 67
6.2.2 - The Silsoe Structures Building. .. 68
6.2.3 - The Silsoe Experimental Wall. .. 69
6.3 - CFD test Accuracy Checks. .. 70
6.3.1 - Introduction. .. 70
6.3.2 - Use of Full-Scale Results for CFD Validation. .. 70
6.3.3 - CFD Solver .. 72
6.3.4 - Geometry. .. 72
6.3.5 - Atmospheric Boundary Layer Inlet Conditions. .. 74
6.3.5.1 - SRI: Boundary Layer Inlet Conditions for the
k- Turbulence Model. ........................ 74
6.3.5.2 - Appropriate Inlet Conditions for the DSM. .. 77
6.3.6 - Variations on the Ground and Building Roughness. .. 78
6.3.7 - Revision to the Wall Functions for the High Reynolds
Number Turbulence Models. .. 79
6.3.8 - Assessment of the y
+
Values. .. 81
6.3.9 - Grid Refinement Tests. .. 81
6.3.10 - Double Precision Checks. .. 83
6.3.11 - 6m Cube Model Reynolds Number Checks. .. 84
6.3.12 - 6m Cube Model Effect of Differencing Scheme. .. 84
6.3.13 - Use of High Reynolds Number Turbulence Models. .. 85
6.3.14 - Solution Convergence. .. 86
6.3.15 - Calculation of Pressure Coefficients. .. 87
6.3.16 - Accuracy Test Summary. .. 88
6.4 - Turbulence Model Validation The 2D Backstep. .. 89
6.5 - CFD and Full-Scale Test Results. .. 91
6.5.1 - Cube Normal to the Incident Wind. .. 91
6.5.2 - Roof / Wake Flow Field Patterns. ... 92
v
6.5.3 - Normal Cube Mean Pressure Distribution. ... 92
6.5.4 - Turbulent Kinetic Energy Distribution. ... 94
6.5.5 - Pressure and Velocity Distribution: Cube Orientated at 45. ... 95
6.6 - The Silsoe Structures Building. ... 96
6.6.1 - Introduction. ... 96
6.6.2 - Flow Field and Pressure Coefficient Distribution The SSB
with Sharp Eaves / Normal Orientation. 97
6.6.3 - Pressure Coefficient Distribution The SSB with Curved
Eaves / Normal Orientation. ... 99
6.6.4 - Pressure Coefficient Distribution The SSB with Curved
Eaves / Transverse Orientation. ... 100
6.7 - The Silsoe Experimental Wall Structure. ... 102
6.7.1 - Introduction. ... 102
6.7.2 - Pressure Coefficient Distribution: Experimental Wall at 0
to the incident wind. ... 102
6.7.3 - Pressure Coefficient Distribution: Experimental Wall at 30
to the incident wind. ... 103
6.7.4 - Pressure Coefficient Distribution: Experimental Wall at 60
to the incident wind. ... 104
6.8 - Turbulence Model: Discussion ... 104
6.8.1 The Standard k- Turbulence Model. ... 104
6.8.2 - The RNG k- Turbulence Model. ... 106
6.8.3 - The MMK k- Turbulence Model. ... 108
6.8.4 - The Quadratic Non-Linear k- Turbulence Model. ... 108
6.8.5 - The Differential Stress Turbulence Model. ... 110
6.8.6 - Results Summary. ... 111
7. Large Eddy Simulation
7.1 Introduction. 113
7.2 - Theory : Filtering the Navier-Stokes Equations. 113
7.3 - Sub-Grid Scale Models. 115
7.3.1 - The Smagorinsky Model. 116
7.4 - Numerical Methods The Time Dependent Term. 118
7.5 - Model Validation and Analysis Detail. 118
7.5.1 - Domain Size. 119
7.5.2 - Inlet Conditions. 120
7.5.3 - Determination of Time-Step. 121
7.5.4 - Large Eddy Turnover Time. 121
vi
7.5.5 - Wall Conditions. 121
7.5.6 - Calculation of Velocity Statistics. 123
7.6 - Results. 123
7.6.1 - Introduction. 123
7.6.2 - Steady State Simulation and LES Results. 124
7.6.3 - Discussion and Recommendations for Future Work 126
8. Conclusions and Recommendations for Future Work.
8.1 - Introduction. 128
8.2 - Conclusions. ... 128
8.3 - Recommendations for Future work. .... 131
Nomenclature. 133
References. 135
Appendix A
Figure numbers 1 to 131 155
vii
Abstract.
The fundamental errors in the numerical modelling of the turbulent component of
fluid flow are one of the main reasons why computational fluid dynamics techniques
have not yet been fully accepted by the wind engineering community.
This thesis is the result of extensive research that was undertaken to assess the various
methods available for numerical simulation of turbulent fluid flow. The research was
undertaken with a view to developing improved turbulence models for computational
wind engineering. Investigations have concentrated on analysing the accuracy and
numerical stability of a number of different turbulence models including both the
widely available models and state of the art techniques.
These investigations suggest that a turbulence model, suitable for wind engineering
applications, should be able to model the anisotropy of turbulent flow as in the
differential stress model whilst maintaining the ease of use and computational
stability of the two equation k- models. Therefore, non-linear expansions of the
Boussinesq hypotheses, the quadratic and cubic non-linear k- models, have been
tested in an attempt to account for anisotropic turbulence and curvature related strain
effects.
Furthermore, large eddy simulations using the standard Smagorinsky sub-grid scale
model have been completed, in order to account for the four dimensional nature of
turbulent flow. This technique, which relies less heavily on the need to model
turbulence by utilising advances in computer technology and processing power to
directly resolve more of the flow field, is now becoming increasingly popular in the
engineering community.
The author has detailed and tested all of the above mentioned techniques and given
recommendations for both the short and longer term future of turbulence modelling in
computational wind engineering.
Improved turbulence models that will more accurately predict bluff body flow fields
and that are numerically stable for complex geometries are of paramount importance
if the use of CFD techniques are to gain wide acceptance by the wind engineering
community.
viii
List of figures.
- Figure 1 Details of the atmospheric boundary layer. .. 155
- Figure 2 The effect of a uniform wind profile on a surface mounted
cube. .. 155
- Figure 3 The effect of a boundary layer mean wind profile on a
surface mounted cube. .. 156
- Figure 4 The flow at the side walls. .. 156
- Figure 5 The flow over the roof of a surface mounted cube with
uniform incident wind. .. 157
- Figure 6 The flow over the roof of a cube with an ABL incident wind. .. 157
- Figure 7 Cube skewed at 45 degrees to the ABL incident wind. .. 158
- Figure 8 Flow in the wake. .. 158
- Figure 9 Reynolds stresses: convention. .. 159
- Figure 10 Finite differences. .. 159
- Figure 11 Control volume notation. .. 160
- Figure 12 Overshoots using the QUICK method. .. 160
- Figure 13 A comparison of CCCT and QUICK interpolation profiles. .. 160
- Figure 14 The 6m cube (site photograph). .. 161
- Figure 15 The Silsoe Structures Building (site photograph). .. 161
- Figure 16 The Silsoe Experimental wall (site photograph). .. 162
- Figure 17 Position of the outlet boundary. .. 162
- Figure 18 Comparison of boundary layer velocity profiles for the
DSM and standard k- models. .. 163
- Figure 19 Comparison of boundary layer turbulent kinetic energy
profiles for the DSM and standard k- models. .. 163
- Figure 20 6m cube, normal orientation, comparison of windward
face pressure distributions for different boundary layer
and building roughness heights. .. 164
- Figure 21 6m cube, normal orientation, comparison of roof pressure
distributions for different boundary layer roughness heights
and building roughness. .. 164
- Figure 22 6m cube, normal orientation, comparison of windward face
pressure distributions for the CFX original and revised wall
functions. 165
- Figure 23 6m cube, normal orientation, comparison of roof pressure
distributions for the CFX original and revised wall functions. .. 165
- Figure 24 6m cube, normal orientation, comparisons of windward face
ix
pressure distributions for the grid refinement test number 1. ... 166
- Figure 25 6m cube, normal orientation, comparisons of roof pressure
distributions for the grid refinement test number 1. .. 166
- Figure 26 6m cube, normal orientation, comparisons of roof pressure
distributions for the grid refinement test number 2. .. 167
- Figure 27 6m cube, normal orientation, comparisons of roof pressure
distributions for the grid refinement test number 2. .. 167
- Figure 28 6m cube, skewed orientation, comparisons of roof pressure
distributions for the grid refinement tests. ..... 168
- Figure 29 6m cube, skewed orientation, comparisons of roof pressure
distributions for the grid refinement tests. . 168
- Figure 30 6m cube, skewed orientation, comparisons of the roof lateral
velocity component for the grid refinement tests. .. 169
- Figure 31 SSB, sharp eaves, windward roof pressure distribution, grid
refinement test. .. 169
- Figure 32 Silsoe experimental wall, windward face pressure distribution,
grid refinement test. .. 170
- Figure 33 Convergence history, residual plot to show solution
convergence for a single precision run. .. 170
- Figure 34 Convergence history, residual plot to show solution
convergence for a double precision run. .. 171
- Figure 35 6m cube, normal orientation, roof pressure distribution,
Reynolds number test. .. 171
- Figure 36 6m cube, normal orientation, roof vortex strength, Reynolds
number test. .. 172
- Figure 37 6m cube, normal orientation, roof pressure distribution,
comparison of differencing schemes. . 172
- Figure 38 6m cube, normal orientation, roof turbulent kinetic energy
distribution, comparison of differencing schemes. . 173
- Figure 39 6m cube, normal orientation, roof vortex strength,
comparison of differencing schemes. .... 173
- Figure 40 False diffusion with the hybrid differencing scheme. .... 174
- Figure 41 False diffusion with the CCCT differencing scheme. .. 174
- Figure 42 The computational domain for the 6m cube model. .. 175
- Figure 43 The CFD domain for the 6m cube simulations. .. 175
- Figure 44 Example of the CFD mesh for the grid refined simulations. .. 176
- Figure 45 Normal cube measurement locations. .. 176
- Figure 46 6m cube, normal orientation, comparison of roof vortex strength
x
and reattachment point. ..... 177
- Figure 47 6m cube, normal orientation, comparison of wake vortex strength
and reattachment point. ........ 177
- Figure 48 Velocity vectors showing the roof flow field for the 6m cube,
standard k- model. .... 178
- Figure 49 Velocity vectors showing the roof flow field for the 6m cube,
MMK k- model. 178
- Figure 50 Velocity vectors showing the roof flow field for the 6m cube,
RNG k- model. 179
- Figure 51 Velocity vectors showing the roof flow field for the 6m cube,
non-linear k- model. 179
- Figure 52 Velocity vectors showing the roof flow field for the 6m cube,
differential stress model. 180
- Figure 53 6m cube, normal orientation, windward face pressure
distribution. 180
- Figure 54 6m cube, normal orientation, roof pressure distribution. 181
- Figure 55 6m cube, normal orientation, leeward face pressure
distribution. 181
- Figure 56 6m cube, normal orientation, side face pressure
distribution. 182
- Figure 57 6m cube, normal orientation, lateral roof pressure
distribution. 182
- Figure 58 6m cube, normal orientation, windward face turbulent kinetic
energy distribution. 183
- Figure 59 6m cube, normal orientation, roof turbulent kinetic energy
distribution. 183
- Figure 60 6m cube, normal orientation, centreline turbulent kinetic energy
contours for the standard k- model. 184
- Figure 61 6m cube, normal orientation, centreline turbulent kinetic energy
contours for the MMK k- model. 184
- Figure 62 6m cube, normal orientation, centreline turbulent kinetic energy
contours for the RNG k- model. 185
- Figure 63 6m cube, normal orientation, centreline turbulent kinetic energy
contours for the non-linear k- model. 185
- Figure 64 6m cube, normal orientation, centreline turbulent kinetic energy
contours for the differential stress model. 186
- Figure 65 Skewed cube measurement locations. .. 186
xi
- Figure 66 6m cube, skewed orientation, comparison of roof lateral
velocity component. . 187
- Figure 67 Skewed cube velocity vectors showing the delta wing
vortices using the non-linear k- model. . 187
- Figure 68 6m cube, skewed orientation, comparison of windward
face pressure coefficients (measurement line 1). . 188
- Figure 69 6m cube, skewed orientation, comparison of roof pressure
coefficients (measurement line 2). . 188
- Figure 70 6m cube, skewed orientation, comparison of roof pressure
coefficients, grid refined results. . 189
- Figure 71 6m cube, skewed orientation, comparison of leeward face
pressure coefficients (measurement line 3). . 189
- Figure 72 Schematic of the SSB / measurement line locations .. 190
- Figure 73 SSB, sharp eaves, normal orientation, windward face
pressure (measurement line 3). . 190
- Figure 74 SSB, sharp eaves, normal orientation, windward face
turbulent kinetic energy (measurement line 3). . 191
- Figure 75 SSB, sharp eaves, normal orientation, windward roof
pressure (measurement line 3). . 191
- Figure 76 SSB, sharp eaves, normal orientation, windward roof
turbulent kinetic energy (measurement line 3). . 192
- Figure 77 SSB, sharp eaves, normal orientation, centreline turbulent
kinetic energy contour plot for the non-linear k- model. . 192
- Figure 78 SSB, sharp eaves, normal orientation, leeward roof
pressure distribution (measurement line 3). . 193
- Figure 79 SSB, sharp eaves, normal orientation, leeward face
pressure distribution (measurement line 3). . 193
- Figure 80 SSB, curved eaves, normal orientation, windward face
pressure distribution (measurement line 3). . 194
- Figure 81 SSB, curved eaves, normal orientation, windward roof
pressure distribution (measurement line 3). . 194
- Figure 82 SSB, curved eaves, normal orientation, leeward roof
pressure distribution (measurement line 3). . 195
- Figure 83 SSB, curved eaves, normal orientation, leeward roof
kinetic energy distribution (measurement line 3). . 195
- Figure 84 SSB, curved eaves, normal orientation, leeward face
pressure distribution (measurement line 3). . 196
- Figure 85 SSB, curved eaves, transverse orientation, front face
xii
pressure distribution (measurement line 1). ...... 196
- Figure 86 SSB, curved eaves, transverse orientation, roof pressure
distribution (measurement line 1). .. 197
- Figure 87 SSB, curved eaves, transverse orientation, centreline velocity
vector plot. . 197
- Figure 88 SSB, curved eaves, transverse orientation, front face pressure
distribution (measurement line 2). . 198
- Figure 89 SSB, curved eaves, transverse orientation, roof pressure
distribution (measurement line 2). 198
- Figure 90 SSB, curved eaves, transverse orientation, front face pressure
distribution (measurement line 3). 199
- Figure 91 SSB, curved eaves, transverse orientation, roof pressure
distribution (measurement line 3). 199
- Figure 92 Schematic of the modular panel layout of the experimental
wall. . 200
- Figure 93 Silsoe experimental wall, windward face pressure distribution
(0 case). 200
- Figure 94 Silsoe experimental wall, leeward face pressure distribution
(0 case). 201
- Figure 95 Silsoe experimental wall, windward face pressure distribution
(30 case). 201
- Figure 96 Silsoe experimental wall, leeward face pressure distribution
(30 case). 202
- Figure 97 Turbulent kinetic energy contours for the experimental wall,
standard k- model (30 case). 202
- Figure 98 Turbulent kinetic energy contours for the experimental wall,
non-linear k- model (30 case). 203
- Figure 99 Velocity vectors for the experimental wall, standard k- model .
(30 case). 203
- Figure 100 Velocity vectors for the experimental wall, non-linear k- model
(30 case). 204
- Figure 101 Silsoe experimental wall windward face pressure distribution
(60 case). 204
- Figure 102 Silsoe experimental wall leeward face pressure distribution
(60 case). 205
- Figure 103 Velocity vectors for the experimental wall, non-linear k- model
(60 case). 205
xiii
- Figure 104 6m cube, centreline uw shear stress contour plot, non-linear
k- model. 206
- Figure 105 6m cube, centreline uw shear stress contour plot, DSM. .... 206
- Figure 106 6m cube, centreline uu normal stress contour plot, non-linear
k- model. 207
- Figure 107 6m cube, centreline vv normal stress contour plot, non-linear
k- model. 207
- Figure 108 6m cube, centreline ww normal stress contour plot, non-linear
k- model. .... 208
- Figure 109 6m cube, centreline uu normal stress contour plot, linear
k- model. 208
- Figure 110 6m cube, centreline vv normal stress contour plot, linear
k- model. 209
- Figure 111 6m cube, centreline ww normal stress contour plot, linear
k- model. .... 209
- Figure 112 Normal Reynolds stress distribution for the roof of the 6m cube
calculated using the non-linear k- model. 210
- Figure 113 Normal Reynolds stress distribution for the roof of the 6m cube
calculated using the differential stress model. 210
- Figure 114 Normal Reynolds stress distribution for the wake of the 6m cube
calculated using the non-linear k- model. 211
- Figure 115 The computational domain for the LES simulations. 211
- Figure 116 LES, time series of fluctuating velocity for the 1/30 scale cube
model. 212
- Figure 117 LES transient velocity vector plots, 1/30 scale cube model,
to 120 coarse grid (plan view). ....... 212
- Figure 121 LES, frequency spectra for the 1/30 scale cube model. ... 214
- Figure 122 LES transient flow, velocity vector plot, fine resolution grid,
1/30 scale cube model (plan view). 215
- Figure 123 LES transient flow, velocity vector plot, 1/30 scale SSB model
(side view). 215
- Figure 124 LES transient flow, velocity vector plot, 1/30 scale SSB model,
full domain (side view). 216
- Figure 125 LES transient flow, velocity vector plot, 1/30 scale cube model,
coarse grid, domain and flow field (plan view). . 216
- Figure 126 LES transient flow, velocity vector plot, coarse grid (side view). .. 217
- Figure 127 LES, time averaged velocity contour plot, 1/30 scale cube model,
coarse grid (side view). 217
xiv
- Figure 128 LES, time averaged pressure contour plot, 1/30 scale cube model,
coarse grid (side view). 218
- Figure 129 Roof velocity vectors for the LES comparisons predicted using
the (steady state) non-linear k- model. 218
- Figure 130 Roof velocity vectors for the LES comparisons predicted using
the (steady state) low Reynolds number k- and k- models. .. 219
- Figure 131 Wake and roof velocity vectors for the LES comparisons predicted
using the (steady state) non-linear k- model. 219
xv
List of tables.
- Table 1 Wind damage costs. 11
- Table 2 CPU requirements for DNS. 61
- Table 3 CPU requirements for CSC. 64
- Table 4 Accuracy test summary. 89
- Table 5 Reattachment length for the backward facing step. 90
- Table 6 Reattachment lengths: cube normal to the incident wind. 92
- Table 7 Reattachment lengths for the front corner vortex: the SSB. 97
- Table 8 Results summary. 112
- Table 9 Roof and wake reattachment distances for LES
and the comparative steady state simulations. 125
xvi
Acknowledgements.
Profound thanks to the following people:
Dr N.G. Wright, for his full support, help and encouragement throughout this study
and his contribution to the making the Ph.D. such a rewarding experience.
Professor Baker, for his continuing help and encouragement.
Dr. Andrew Quinn, a source of great help throughout the Ph.D, and someone whom,
in the early days, had a good deal of CFD knowledge when I had none!.
Dr D. Hargreaves, a very helpful and extremely clever chap.
Dr R. Hoxey, definitely the man to speak to when it comes to experimental fluid
mechanics (and CFD too).
Dr I.P. Jones and N. Wilkes of CFX-International, without whom this Ph.D. would
possibly have been much shorter and not particularly interesting!.
Matt Straw, for many hours of discussion putting (CFD and) the world to rights.
My wife Samantha, for putting up with me, an eternal student, for all these years
while she has gone out and put food on the table.
My family for their continued support and encouragement through the years.
xvii
List of publications.
Wright, N.G., Easom, G.J. and Hoxey, R. (1998). Improved computational models
for wind engineering., Proceedings of the 4
th
UK Conference on Wind Engineering,
Bristol University, Bristol, England.
Wright, N.G., Easom, G.J. and Hoxey, R. (1999). Improved computational models
for wind engineering., Proceedings of the 10
th
ICWE, Copenhagen, Denmark.
Wright, N.G. and Easom, G.J. (1999). Comparison of several computational
turbulence models with full-scale measurements of flows around buildings., Paper
accepted for publication in the Journal of Wind and Structures.
Wright, N.G. and Easom, G.J. (1999). Improved computational models for wind
engineering., Paper accepted for publication in the Journal of Wind and Structures.
1. Introduction.
Generally Computational Fluid Dynamics can be defined as the analysis of
systems involving fluid flow, heat transfer and associated phenomena such as
chemical reactions by computer based simulation. (Versteeg and
Malalasekera 1995, p1)
The majority of practical fluid flows are turbulent and consist of many complex flow
features which may contain, among other features, recirculation zones and flow
stagnation points. The types of flows encountered in the field of wind engineering are
no exception.
Traditionally, due to the complexity of this subject, all the research and design work
undertaken in this field has concentrated on the use of full-scale and wind tunnel
analysis. This has involved the use of expensive wind tunnel and data recording
facilities and has required significant time and effort to obtain the desired results.
However during the 1970s and 1980s there was a great deal of interest among the
engineering community into a relatively new technique known as computational fluid
dynamics (CFD). The advances made in high speed digital computer technology had
enabled the solution of flow problems, which were described mathematically by a set
of coupled nonlinear partial differential equations and the appropriate boundary
conditions, in a relatively short space of time and for a low financial cost. Initially the
wind engineering community largely ignored this technique due to the need for
powerful computers and the errors in early modelling techniques. Nonetheless, the
rapidly falling costs of computer hardware and further advances in technology in the
late 1980s and early 1990s enabled CFD to be applied to the complex field of wind
engineering.
In theory it is numerically possible to completely resolve all aspects of a fluid
dynamics problem including the rapid spatial and temporal variations of turbulence in
the flow using a CFD technique known as direct numerical simulation. This technique
involves discretising the equations using the finite volume method at a mesh size
below the smallest eddies in turbulent flow, the Kolmorogov length scale, and
therefore resolving the flow down to the smallest spatial and temporal variations.
Unfortunately the direct numerical simulation of practical turbulent fluid flows using
the time dependent Navier-Stokes equations in their simplest form is well beyond the
capabilities of present day computing power. This is due to the fact that the amount of
1
2
computer processing (CPU) time required is dependent on the degree of resolution of
the small scale eddies. The smallest eddies in turbulent flow, the so-called
Kolmogorov microscale, are very small at about 0.1 to 1mm for natural wind
(Murakami 1997). Therefore the numerical discretisation of an entire wind
engineering flow field with a complex geometry at high Reynolds numbers is at
present well beyond the capabilities of even the most powerful supercomputers
available.
The only economically feasible way to solve this problem is to employ statistically
averaged equations which govern the mean flow equations. Turbulence models are
then required to achieve closure of the averaged equations and represent the action of
turbulent stresses on the mean flow. Unfortunately the mathematical models used in
CFD are only able to perform as well as the physical assumptions and knowledge
built into them will allow. In particular the assumptions made regarding the modelling
of the turbulent component of engineering flows have proved to be a major source of
error in wind engineering simulations.
Presently, the most popular and widely used models use equations to represent a
single length and velocity scale and are based on Reynolds averaging and the
isotropic eddy viscosity concept. Although many of these turbulence models have
been used successfully in aeronautical applications, in which fluid flow without
separation may be a regular occurrence the same is not true of wind engineering
applications. Wind engineering flow fields are highly complex and are characterised
by the presence of multiple recirculation zones embedded within a uni-directional
flow. The addition of streamline curvature and favourable and adverse pressure
gradients leads to flow fields possessing very different turbulence scales and
structures. Consequently such turbulence models have great difficulty in simulating
wind engineering flow fields which are essentially transient and highly anisotropic.
It is therefore apparent that one of the main obstacles to the use of CFD in wind
engineering is that of turbulence modelling.
In view of these shortcomings the aim of this work has been to conduct research into
the various turbulence modelling methods available with a view to developing
improved turbulence models for computational wind engineering. This project has
concentrated on a number of different turbulence models and analysed their effects on
the accuracy of the results obtained for bluff body flow simulations from the CFD
package CFX. A number of bluff body test cases have been used, all of which have
3
also been analysed experimentally at full-scale to allow direct comparisons of the
available data. As far as the author is aware these tests represent the first comparison
of experimental data from a full-scale surface mounted cube, rather than model scale,
to CFD derived results.
This thesis is structured as follows:
Chapter 2 introduces the reader to the subjects of wind engineering and computational
wind engineering (CWE).
All the important numerical considerations regarding the computational analysis of
fluid flow are reviewed in chapter 3. This chapter offers comprehensive detail on all
the important aspects of numerical discretisation including methods of dealing with
the convection terms in the transport equations.
Chapters 4 and 5 detail the main problem area for attaining accurate CFD simulations,
that of turbulence modelling. Chapter 4 reviews both the more conventional models in
wide spread use today and a number of more theoretical models and highlights their
errors and potential inadequacies. Chapter 5 summarises the main problems with these
conventional models and offers a number of possible improvements.
Chapter 6 details the performance of many of the models described in chapters 4 and
5 for a number of different bluff body flow fields and compares the results obtained
against full-scale data obtained from the Silsoe Research Institute. Full analysis and
discussion of the results is offered based on the ability of the model to meet certain
criteria, described later, in order to assess their applicability as a wind engineering
turbulence model.
Chapter 7 highlights a completely different technique for modelling turbulent flow
fields described as Large Eddy Simulation. This was separated from chapter 6 due to
the fact that this technique is based on a different method of filtering the main fluid
flow equations and different methods have been used to ascertain its accuracy. Due to
the large computational overheads of this technique this chapter was mainly used to
assess whether LES can be applied to flows with high Reynolds numbers and large
flow domains, i.e. wind engineering applications and to assess the future work to be
undertaken. Full analysis and discussion is offered.
4
Results obtained from these investigations have clearly shown the essential
requirements for turbulence models to adequately predict bluff body flow fields as
highlighted in chapter 8.
5
2. Wind Engineering.
2.1 Introduction
This chapter provides a general introduction into the field of wind engineering with a
brief overview of this complex and diverse subject. The reader is offered a brief
history of the subject with reference to past disasters and the more prominent figures
who have contributed to this subject. The significant effects of the atmospheric
boundary layer on the surface pressures around buildings are graphically
demonstrated and a brief discussion of the available wind loading codes of practice
for structural design is given. This chapter then introduces the reader to the following
chapters through a brief discussion of computational wind engineering focusing on
the relative benefits of applying this technique to wind engineering problems.
2.2 Wind Engineering - Bluff Body Aerodynamics
Wind engineering is a relatively young field, with the advent of computational wind
engineering perhaps as little as ten or fifteen years ago. Nevertheless the development
of both computational and experimental wind engineering has been fast, building on
the knowledge gained during the explosive expansion of fluid dynamics at the
beginning of the century, in the rush to perfect and exploit the aeroplane (Cook
1986a). The foundations of wind engineering and fluid mechanics can be traced back
to the late seventeenth century with advances and discoveries by such prominent
figures as Newton, Hook and Bernoulli. The discipline has progressed from these
early years through a combination of advances made in other related disciplines such
as aeronautics and through a series of high profile failures in the early 19
th
and 20
th
centuries. Such disasters include the collapse of the Tay and the Tacoma Narrows
bridges in 1879 and 1940 respectively. These gave added impetus to the need to
understand wind effects in the lower atmosphere and led to pioneering experiments
being undertaken by Reynolds, Stokes and Jenson. Unfortunately until about 1960 it
was still assumed that the general techniques applied to aeronautics could be
successfully applied to ground based structures. As such wind forces on buildings
were determined using relatively smooth inflow conditions in the various wind tunnel
experiments undertaken. Experiments that compared wind tunnel results of models in
smooth air to full-scale studies discovered major discrepancies that were quickly
recognised and explained. These events and discoveries have led to a sound
6
understanding of wind effects in the lower atmosphere and have brought about the
development of design codes that allow safer structures to be built.
2.2.1 Fundamental Aspects
Wind engineering is concerned with the accurate calculation of wind induced loading
of a structure and the ability of the said structure to resist such forces without failure.
Consequently there are three fundamental considerations as follows:
1. the wind climate comprising the weather
2. the atmospheric boundary layer
3. the structure.
The first point is important as various locations in the world experience vastly
different extreme weather systems due to their respective climates i.e. polar, tropical
and moderate. In the United Kingdom, the climate is dominated by prevailing
westerly winds and by a band of low pressure between Scotland and Ireland. Atlantic
depressions caused by the instability of the polar front track across the Atlantic and
dominate the weather systems in the UK (Cook 1986a). Consequently as these
depressions pass over the UK they influence different parts of the country in varying
degrees of severity.
The Atmospheric Boundary Layer (ABL) is the result of the interaction of the wind
and the Earths surface. Between the Earths surface and the top of the boundary
layer, the gradient height, the wind is retarded due to roughness elements such as
grass, trees, buildings and general relief or topography. These effects are transmitted
upwards by Reynolds stresses resulting from the exchange of momentum between
successive layers of the atmosphere. The ground level shear stresses increase with
increasing ground roughness (Cook 1986a). The overall effect of the ABL on the
incident wind at ground level is one of the main factors that makes the field of wind
engineering so different from, for example, aeronautical engineering. The turbulence
generated by the retarding effect of the ground results in a significant increase in the
complexity of the incident wind and the effects on ground based structures. The ABL
can be further divided into a number of sub-layers as follows:
Roughness Sub-layer This layer extends from the surface to the average height of
the roughness elements that may range from snow and grass to trees and buildings.
7
Inertial Sub-layer This layer extends from the roughness layer to a height that is
dependent on the degree of roughness of the surface retarding the flow. Both the
roughness and inertial sub-layers are encompassed within the surface layer.
Ekman layer extends from the top of the surface layer to a height where the wind is
unaffected by the Earths surface. This height is commonly referred to as the gradient
height.
A general idea of the extent of these regions is given in Figure 1.
Further considerations that add to the complexity of the incident wind at any location
are apparent due to the surrounding topography and fetch. Sites located near to hills
and inclines may experience higher winds or be sheltered depending on the particular
location and wind direction. The development of the boundary layer and roughness of
the fetch has to be taken into consideration in determining the possible wind speeds,
turbulence levels and therefore gust strengths that affect a structure at a given
location. The affect of altitude and seasonal variations in weather also needs to be
included.
The method of construction and shape of the structure define the way in which the
atmospheric boundary layer interacts with a building. Considering the building as a
whole, when a structure is small it will be loaded by small gusts and the full range of
frequencies in the boundary layer will be significant. Conversely when the structure is
large the smaller gusts will not act simultaneously on the structure and will tend to
cancel each other out so that only the lower frequencies are significant. When a
structure is stiff it will have a high natural frequency in each of its first few nodes and
will tend to follow the fluctuations of load without significant amplification or
attenuation (Cook 1986a). As the lowest modal frequencies are high there is little
energy in the spectrum of the atmospheric turbulence available to excite resonance.
Deflection of the structure is not significant and as such the structure is said to be
static. This class of structure represents the majority of buildings in wind engineering
analysis. The only design parameter of importance for a static structure is the
maximum load likely to be experienced in its lifetime (Cook 1986a). The designer
therefore ensures that the design strength of the structure exceeds this load by a
suitable safety margin to account for variations of material strengths and uncertainties
in the assessment. Other structures, which are susceptible to deflection, may come
under the heading of dynamic or even aero-elastic structures. These can be excited by
8
loading at frequencies at the natural frequency of the structure, which may result in
excessive deflections. Therefore the excessive deflection and the consequent increase
in stresses will become important in designing the structure.
The majority of the damage that occurs with static structures in the UK follows major
windstorms such as the 1987 storm that affected the south of England. Damage is
usually concentrated on the cladding or roofs of buildings in localised areas of high
negative (suction) pressures, as detailed in the following section. In addition,
depending on the direction of the incident wind to the structure, delta wing vortices
may occur resulting in very high negative pressures. Fatigue loading of cladding may
also occur as a consequence of repeated or fluctuating loading on a surface which can
result in the general weakening of structural ties and connections. Further details can
be found in Cook (1986b).
2.2.2 Bluff Body Flow Fields
Wind engineering flow fields are complicated firstly by virtue of the turbulent
atmospheric boundary layer, as briefly detailed, and secondly due to the fact that we
are mainly dealing with non-aerodynamic or bluff bodies, defined as follows:
A body is aerodynamically bluff when the flow streamlines do not follow the
surface of the body, but detach from it leaving regions of separated flow and a
wide trailing wake. (Cook 1986a, p27)
The flow fields around a classical bluff body, a surface mounted cube will now be
briefly described in order to highlight both the effects of the ABL and the complex
flow fields produced. It is worth noting that the descriptions to follow are simplistic
and that there are complex interactions between the different flow effects. Castro and
Robbins (1977) give a full and detailed description of the following.
For the windward face in a uniform wind the maximum pressure is located on the
centreline of the face next to the ground as all kinetic energy of the flow is transferred
to dynamic pressure as the flow velocity reduces to zero (see Figure 2).
The revised incident wind gradient, best represented by a logarithmic law profile (see
Figure 3) allows the flow to move down the face in addition to the tendency to move
around the sides and over the top of the cube. Above the front stagnation point the
9
flow rises over the top of the building resulting in a lower mass of air flowing over the
roof. Below this point the flow travels down the front face of the cube until it reaches
the ground level. As the flow moving down the front of the cube has more kinetic
energy than the incident wind at this level it is able to move upstream against the
incident wind. Eventually all the kinetic energy of the flow is exhausted in working
against the incident flow and the vortex is formed in front of the building as the air
rolls up next to the ground. The existence of this vortex has a marked effect on the
distribution of pressure on the windward face. This frontal vortex facilitates formation
of the horseshoe vortex at the sides of the building (Cook 1986a).
The existence of the frontal vortex, due to the ABL incident flow, has the effect of
bringing faster airflow to ground from higher up the front face of the cube (see Figure
4). Air entering this vortex escapes around the sides of the cube resulting in
significantly faster air flow at near ground level than found in the incident wind at this
level. This results in a vortex forming around the front and sides of the cube which
has a horseshoe shape and is thus termed a horseshoe vortex. This in turn causes high
suction pressures at the upwind edge of the side face and high values of lateral
diffusion in the wake of the building (Castro and Robbins 1977).
For the roof of the building with a uniform incident wind the flow from the full height
of the building rises over the roof and does not reattach (see Figure 5).
For the boundary layer incident wind case, only the flow from the upper third of the
cube rises over the roof. This effect combined with the increased mixing effect of the
higher levels of turbulence tends to make the flow reattach at some point along the
roofline (see Figure 6).
For a cube skewed to the incident wind the flow separating at the upwind corner will
tend to be displaced under the flow separating immediately downwind of the corner.
The net effect is that the vorticity of the flow is increased until a strong conical vortex
is formed (shown in Figure 7) resulting in very high negative pressures. If the cube is
skewed to the flow by 45 then a symmetrical pair of vortices will occur. If this is not
the case then the vortices will be of different strengths.
The flow in the wake of a surface mounted cube is defined by flow recirculation (see
Figure 8) along with unsteady fluctuating pressures and Karmans vortex street.
10
2.2.3 Codes of Practice
The structural designer or wind engineer must be able to produce designs that will
safely withstand the wind loads produced by the prevailing weather systems and by
the effects of the ABL and surrounding topography. At present there are two wind
loading codes of practice available in the UK, as will be briefly detailed in the
following:
CP3 Chapter V part II Codes of Basic Data for the Design of Buildings 1972.
This code is generally accepted as being one of the first wind loading codes of
practice available and was completed prior to major advances in the understanding of
the atmospheric wind. As such the loading model employed by this code is a
simplified version of the earliest available model for the action of turbulent wind, the
quasi-steady technique simplified to the equivalent steady gust model (see Cook
(1986a) and (1986b) for details). Very briefly this model assumes that the crosswind
and vertical turbulence terms can be disregarded. This leads to the modelling of the
atmospheric turbulence as temporary variations of incident wind speed with no
changes of azimuth or elevation angle (Cook 1986a). Therefore the code tends to mix
conservatism in some sections of the wind speed estimation along with significant
under estimation of the effects of topographical features. Furthermore, the code does
not take account of the effects of upwind terrain changes or fetch which have marked
effects on wind speeds. Finally Eaton and Newbury (1972) state that, due to the
increasing complexity of structures, this code no longer covers the full range of
building designs that are now available.
BS6399 part 2, 1997 Code of Practice for Wind Loads.
This code represents the most recent method of determining wind speeds for any
given site in the UK. BS6399 is generally superior to the code it supersedes CP3
Chapter V part II, as it accounts for the upwind fetch and fetch changes, topographical
features, directional wind speeds and a far greater range of building shapes and
designs. Furthermore, the wind speed data used in this code was prepared by using the
maximum wind speed during every period of windy weather as opposed to the
analysis of a series of annual maximum wind speeds. Consequently the amount of
11
available data was greatly increased which in turn allowed directional and seasonal
characteristics of the UK wind climate to be examined (BS6399 1997).
2.2.4 Wind Damage Costs
In order to highlight the importance and need for greater understanding of the subject
of wind engineering it is worth briefly discussing the financial costs that can be
incurred as the result of wind storms (see Table 1).
Windstorm event Country Cost to insurers
Million
Winter Storm (1987) Great Britain 1100
Hurricane Gilbert (1988) Jamaica 417
Hurricane Hugo (1989) USA 2617
Winter storms (1990) Germany 1035
Great Britain 2370
Netherlands 600
Table 1: Wind damage costs (Windstorm 1990)
The fact that the 1990 winter storms in the UK were almost as costly as the damage
caused by Hurricane Hugo in the USA gives an indication of the scale of the wind
induced damage that may occur. More importantly consideration must be given to the
injuries and deaths caused by the effects of the wind and structural damage which are
not shown here. In addition, data accumulated over more than 30 years has shown that
the incidence of high winds and wind storms in Britain and around the world has
increased significantly in the last 10 years and is continuing to do so (Windstorm
1990). To a certain extent climatic fluctuations are responsible for this increase in the
extremes in the weather, for example the El- Nino phenomenon. Nevertheless it is
becoming evident that human intervention in the balance of nature is causing a
general change of environmental conditions. This has resulted in climatic change and
more extreme weather conditions and thus increased wind induced structural damage,
injury and loss of life.
2.3 Computational Wind Engineering (CWE)
The preceding sections have briefly demonstrated the complexity of the atmospheric
wind and the ensuing financial and human costs incurred from wind induced failures.
These facts, therefore, highlight the need to be able to accurately assess the effects of
12
the atmospheric wind on ground based structures and the requirement for further
research in this field.
In order to improve our understanding of how the ABL interacts with complex ground
based shapes engineers and scientists have for many years studied the results of full-
scale and wind tunnel tests, the results of which have increased our knowledge of
wind engineering to its present level. Although much is known about the interaction
of wind and surface based structures there is obviously still a good deal to learn.
While it is apparent that the traditional methods will dominate for many years to
come, it is also clear that new technology and techniques should be used whenever
possible to aid the advancement of knowledge. Traditionally, as is the case in wind
engineering, research in structural analysis and design has centered on experiments in
the laboratory. The advent of the electric computer, arguably one of the most
important discoveries of the 20
th
century, has brought about a revolution in this field
allowing quick, accurate and low cost analysis of structural elements. This new
technique is well proven in structural analysis and has now become a major design
tool comparable to, if not exceeding, the use of full-scale experimental apparatus.
Since the first application of CFD to wind engineering approximately 15 years ago the
overall progress in CWE has been a good deal slower. This is mainly due to a
combination of the need for large computing power and a poor understanding of the
physics of fluid flow, namely the phenomena known as turbulence and how it can be
modelled. These problems associated with the numerical solution of fluid flow are
unfortunately further highlighted when we apply this technique to wind engineering
flows due to the highly turbulent nature of the ABL, the non-aerodynamic bluff
bodies, as discussed, and the need for large computational domains for external flow
fields.
2.3.1 The Beginnings of CWE 1980s and early 1990s
The very first use of CFD in wind engineering occurred in the mid to late 1980s by
Summers et al (1986), Matthews (1987) and Murakami and Mochida (1988) with the
application of the standard k- model to flows around building shapes. Early attempts
at large eddy simulation in wind engineering (see section 5.7 and chapter 7) were also
undertaken as early as 1987 by Murakami et al (1987). These tests were in fact the
first to fully analyse the results of a simulation that involved flow impingement and
thus showed the fundamental flaws in the standard k- model based on the isotropic
13
eddy viscosity model, as described in section 4.2.3. This early discovery initiated
one of the most attractive research targets since the start of CFD that of improved
turbulence modelling. The main areas of interest included the following:
1. the improvement of the standard k- model and the introduction of a more
sophisticated Reynolds Averaged Navier-Stokes (RANS) models
2. the development of easier methods for applying LES to CWE problems
(Murakami and Mochida 1999).
The results of this intense research effort to improve the application of CFD to wind
engineering, or to engineering applications in general, have included a number of new
turbulence models. These range from ad-hoc modifications to the k- model to more
advanced RANS models such as the differential stress model, many of which have
been fully tested as part of this project. In addition new large eddy simulation
techniques have allowed for significant improvements in the predictive accuracy of
CFD when applied to wind engineering. Invaluable work over the past 15 years by
researchers and academics such as Leschziner, Speziale, Launder, Rodi and
Murakami, from many different engineering disciplines, have significantly
contributed to improving the accuracy and applicability of CFD techniques.
Unfortunately, although CWE has progressed a long way from the early days of the
1980s, there is still a long way to go before the same confidence can be placed on
CWE as with finite element analysis in structural engineering. There are still many
problems in CWE, including errors in wall boundary conditions and near wall
functions, and although much has been achieved the biggest problem is still that of
turbulence modelling in the highly turbulent and complex flow fields encountered.
Nevertheless, even at the current stage of development, the theoretical advantages of
virtual computer based full scale tests such as the ones undertaken in this project
are significant. So much so that according to Murakami and Mochida (1999) CWE
has evolved into a powerful tool for analysing wind engineering flow fields.
Generally the benefits brought about by the use of CFD in other engineering
disciplines have proved a strong incentive to improve the performance of CFD in
wind engineering. The eventual aim of all CFD wind engineering researchers it to be
able to exactly match a computational solution to the real situation. While it is unclear
at this time whether or not this goal will ever be achieved, it is apparent that continued
14
research is required to, at the very least, attain success in CWE comparable to that
found in, say, aeronautical engineering where CFD is routinely used. As such the
author is of the opinion that the work detailed in the following chapters is very useful
to the wind engineering community in attempting to offer possible improvements to
the problem of turbulence modelling and therefore advancing the use of CFD in wind
engineering.
15
3. Solution Procedures and Numerical Methods.
3.1 Introduction
The aim of this chapter is to introduce the reader to the methods and concept of
numerical analysis of fluid flow. Firstly, there is a brief introduction to the equations
of fluid motion and the simplifications and techniques used to account for turbulence.
This is followed by a description of the discretisation schemes whereby the governing
differential equations are transformed into their corresponding numerical analogues.
A number of common differencing schemes for the convection terms and the
associated numerical problems are also reviewed.
3.2 The Navier-Stokes Equations
The numerical solution of any fluid flow problem requires the solution of the general
equations of fluid motion, the Navier-Stokes and the continuity equations. Fluid flow
problems are described mathematically by these equations which are a set of coupled
non-linear partial differential equations with appropriate boundary conditions. These
equations are derived from Newtons Second Law and describe the conservation of
momentum in the flow.
The general form of the three dimensional incompressible instantaneous Navier-
Stokes equations is as follows, in Cartesian tensor form:
F +
x
u
+
x
) (
=
) (
1
1
]
1

,
_

+
i
j
j
i
j i j
j i
i
x
u
x
P
x
u u
t
u


(3.1)
acceleration convection pressure effects of viscosity body force
term term gradient

and the continuity equation:

t
u
x
i
i
+ = 0 (3.2)
Refer to the nomenclature page for details of the individual terms.
16
For full details of the derivation of these equations the interested reader is referred to
Young (1989).
For low speed laminar flows without heat transfer the equations detailed above can be
used to describe the flow exactly. However, in turbulent flows the velocity
components vary rapidly in both time and space and difficulties arise in the numerical
discretisation of the flow field as briefly described in chapter 1. Furthermore, there is
a major problem in simply representing or modelling turbulence. This is due to the
fact that turbulence is an extremely complex and little understood phenomena which
is defined by a number of highly complex mechanisms including irregularity,
diffusivity, three dimensional vorticity fluctuations and dissipation. A brief
introduction to turbulence theory follows; for a detailed description the interested
reader is referred to Tennekes and Lumley (1972).
Experiments conducted on fluid flow have shown that for Reynolds numbers below a
critical value the flow is found to be smooth and ordered as adjacent layers of the
fluid slide past each other, the flow is said to be laminar and the viscosity of the fluid
dominates. Above the value of the critical Reynolds number the flow state changes to
include more or less random fluctuations superimposed on the mean flow direction
and the flow effectively becomes turbulent as the momentum of the fluid increases.
The length scales or eddy sizes in a turbulent flow are bounded by the dimensions of
the flow field and by the diffusive action of molecular viscosity. Therefore turbulent
flows are characterised by many different velocity and length scales. The largest
eddies do most of the work in the flow field by transferring momentum and
contaminants at rates up to several orders of magnitude greater than that by molecular
diffusion (Launder 1972). These large turbulent eddies interact with and extract
energy from the mean flow by a process called vortex stretching. In turn eddies of a
smaller size are stretched by larger eddies and less by the action of the mean flow.
This process continues with smaller and smaller eddies until all the energy extracted
from the mean flow by the largest eddies is handed down to the small scale dissipative
eddies. This process, where the kinetic energy is effectively passed down from the
largest to the smallest eddies, is termed the energy cascade.
The effect of viscosity dictates the scale of the smallest eddies as an equilibrium is
reached whereby the ability of the eddy to perform work against the action of
viscosity is matched by the dissipative action of the fluid viscosity. The energy of the
rotational eddies is therefore dissipated into thermal energy. These small scale eddies
17
are of the Kolmogorov length. It is believed that the structure of the small scale eddies
at high Reynolds numbers, away from the near wall region, is isotropic as the
diffusive action of viscosity tends to smear out directionality. This is contrasted to the
largest scale eddies which, due to the fact that they interact strongly with the mean
flow, are believed to be highly anisotropic in nature (Tennekes and Lumley 1972).
The most accurate way to model fluid flow numerically is direct numerical simulation
(Murakami 1997) (see section 5.6) which involves discretising the equations at a
mesh size below the Kolmogorov length scale and applying Equation 3.1 along with
suitable wall boundary conditions to the whole flow field. Unfortunately for practical
wind engineering flows this is well beyond the capabilities of present day computers.
Therefore to reduce the amount of computational effort the effect of turbulence has to
be modelled. The starting point in this modelling process is to make the assumption
that the velocity at a given point in space and time can be made up of the
superposition of some mean velocity which varies slowly with time and a random
component that varies rapidly (Shaw 1989).
Therefore the instantaneous velocity component u can be described as:
u + u u = (3.3)
Refer to the nomenclature page for details of the individual terms.
To incorporate the effects of turbulence on the mean flow, Equation 3.3 is substituted
into Equations 3.2 and 3.1 and integrated over time to obtain the mean flow equations.
As the fluctuating components are random and do not show any preferential direction
the integrals over time will be zero for the linear terms in the continuity and
momentum equations. The convective terms in the momentum equation are in fact
non-linear being the product of velocity and derivatives of the velocity component. As
such the convective terms generate extra higher order terms for the products of
fluctuating components. These terms are referred to as the Reynolds stresses.
3.3 The Reynolds Stresses
Proceeding with the averaging process results in the instantaneous values being
replaced with the mean variables except for the case of the convection transport term.
18
Referring to the convection term, (in Equation 3.1) the substitution of the fluctuating
component of velocity results in:
( )( )
i j i j
i j i j i j i j
i i j j i j
u u u u
u u u u u u u u
u u u u u u



+
+ + +
+ +
(3.4)
(Launder 1994)
where the over-bar indicates mean values per unit time.
Thus the time averaged equation of fluid motion becomes:
i j i
i
j
j
i
j i j
i j
i
F u u
u
x
u
x x
u u
t
u
+
x
+
x
+
) (
=
) (
1
1
]
1

,
_


(3.5)
(Launder 1994)
Equation 3.5 is generally referred to as the Reynolds equation and differs from the
equation describing a laminar flow only by the presence of the term containing
averaged products of fluctuating velocity. The process it represents is the additional
transfer of momentum due to turbulent fluctuations. The first term in the brackets is
the viscous term and the second term u u
i j
is the turbulent stress or the Reynolds
stress tensor.

)
)

pairs) l symmetrica (3 = =
stresses shear Three - = =
= =
=
stresses normal Three - =
= ere wh
=

yx
2

2

2
w v
w u
v u
w
v
u
w w v w u w
w v v v u v
w u v u u u
u u
zy yz
zx xz
xy
zz
yy
xx
j i






1
1
1
]
1










19
The convention used to describe the Reynolds stresses is detailed on Figure 9.
In turbulent flows the normal stresses are always non-zero because they contain
squared velocity fluctuations. The shear stresses are associated with correlations
between different velocity components.
It is therefore the main aim of the turbulence model to predict the effect of these
Reynolds stresses on the mean flow. Consequently the next step in the turbulence
modelling process is the formulation and application of a suitable model that can
accurately represent these stresses over a range of flow fields, see chapter 4 for further
details.
3.4 Common Discretisation Schemes
As digital computers are only able to process digital data the partial differential
equations that describe fluid flow, detailed in the previous sections, need to be
transformed into an appropriate form. This process of transforming a partial
differential equation to a numerical analogue of the equation is called discretisation.
There are a variety of techniques available to do this as follows:
1. the finite difference method
2. the finite element method
3. the finite volume method
4. spectral methods.
These methods are described in detail by Versteeg and Malalasekera (1996), Smith
(1978), Zienkiewicz and Taylor (1989) and Patankar (1980). The first three of these
methods are generally similar, modified for application to certain classifications of
partial differential equations (see Smith 1985) used in different engineering fields.
The finite volume method is in fact the most popular discretisation technique used in
CFD applications and is a mixture of the finite difference and finite element
techniques. The basic theory behind these techniques is to build a tool kit of equations
that describe the derivatives of a variable as the difference between successive values
of the variable at various points, elements or volumes in space or time. For example,
for the finite difference method:
20
Using a truncated Taylor series expansion, the value of U at the point x + h and x - h
is equal to (see Figure 10):
U x h U x h
u
x
h
U
x
h U
x
( ) ( ) + + +

_
,

_
,
+

2
2
2
3 3
3
2 6
..... (3.6)
U x h U x h
u
x
h
U
x
h U
x
( ) ( ) +

_
,

_
,
+

2
2
2
3 3
3
2 6
...... (3.7)
where;
h = small displacement.
(Shaw 1989)
Obviously by manipulating these equations terms can be produced for the first and
second derivatives of the function U purely in terms of the x and h as follows:
( ) ( ) ( )
2
2
2
2
) O(h 2
h
h x U x U h x U
x
U + + +

(3.8)
(Shaw 1989)
There are many techniques to obtaining the differencing formula. The method above
would be classed as a central difference formula. Additionally, the differences can be
assessed with reference to a known value prior to or after the value to be assessed, in
terms of positive displacement h. These methods are thus termed backward and
forward differencing.
The finite volume method was originally developed as a special finite difference
formulation. Using this method the partial differential equation representing the
conservation principle for a flow variable over an infinitesimal control volume is
discretised to simply express the same principle over a finite control volume (Patankar
1980). As such the domain of interest is divided into a number of non-overlapping
control volumes such that one control volume surrounds each nodal point. The
boundaries of the control volumes are usually placed midway between adjacent nodes.
Physical and control volume boundaries are also matched. The procedure that
distinguishes the control volume method from the finite difference technique is the
control volume integration. The differential equation expressing the variation of a
flow variable is integrated over each control volume. The resulting discretised
equation expresses the exact conservation of the relevant properties for each finite cell
21
size. For example the conservation of a general flow variable within a finite control
volume can be expressed as a balance between the various processes tending to
increase or decrease it:
Rate of change of flow variable = net flux of variable due + net flux of variable
in control volume w.r.t. time to convection into control due to diffusion
volume
+ net rate of creation of variable inside the
control volume.
(Versteeg and Malalasekera 1996)
Values of the properties of flow variables are defined and evaluated at the nodal
points. To calculate gradients and hence fluxes at the control volume faces an
approximate distribution of properties between nodal points is used. Therefore
interpolated values of the diffusion coefficient at the control volume faces as well as
diffusive flux terms and source terms are substituted into the integral form of the
partial differential equation expressing the flow field for each of the control volumes.
This is repeated for all nodal points in the domain with modifications for control
volumes adjacent to the domain boundaries. The result of this procedure is a system
of linear algebraic equations which are solved through iterative solution procedures to
obtain the distribution of the flow property at all nodal points (see section 3.6.1)
(Versteeg and Malalasekera 1996).
Finally, spectral methods approximate the unknowns by means of truncated Fourier
series. Unlike the above-mentioned approaches the approximations are not local but
valid throughout the computational domain. As a consequence it is reported that this
method is not used in CFD applications as there are difficulties in applying the
technique to flows with complex geometries (Versteeg and Malalasekera 1996). The
interested reader should refer to Gotlieb and Orszag (1977) if further information is
required.
22
3.5 Common Differencing Schemes - Discretisation of the Convection Term
Disregarding the continuity equation, all of the equations solved by CFD packages
have the same general form which includes fluid flow processes such as transient,
convection, advection, diffusion and source terms.
S
x x x
u
t
i i i
i

,
_


. (3.9)
Where is the relevant effective diffusivity for a variable . If an integration of
Equation 3.9 is performed using Gauss law where a volume integral is transformed
into a surface integral the following equation is obtained:

SdV dA
x
dA u V
t
i
i
i I
.

(3.10)
Time dependant. Convection. Diffusion. Source.
The differencing scheme for each term can then be selected to match its physical
characteristics. For example, the diffusion term, which affects the distribution of a
transported quantity along its gradients in all directions, is usually dealt with by
second order centered differencing (see section 3.5.1).
Conversely the convection term in the general equation spreads influence only in the
flow direction as fluid properties are convected downstream. Therefore the size of the
grid cells and the accuracy and properties of the differencing scheme becomes of
rather more importance. As a consequence, differencing schemes which favour
upstream values are preferred.
In theory the numerical results obtained using any differencing scheme should be
indistinguishable from the exact solution for an infinite number of cells. This follows
from the consideration that, as the grid points move closer together the change in the
variables between neighbouring grid points becomes smaller and actual details of the
profile assumption become unimportant. In practice only a finite number of cells are
possible and the differencing schemes have their own unique properties as different
interpolation techniques between neighbouring nodes are employed. Therefore to a
large degree this term determines the accuracy of the solution to the solved equations.
23
In Computational Fluid Dynamics the differencing schemes should have certain
fundamental properties for them to produce physically realistic results. These
properties are known as conservation, boundedness, transportiveness and accuracy.
Conservation
This property ensures global conservation of the fluid properties for the entire
domain. This is a very important property and is achieved by means of consistent
expressions for fluxes through the cell faces of adjacent control volumes in the finite
volume scheme (Versteeg and Malalasekera 1996). This is an important consideration
when different schemes are used to discretise the diffusion and convection terms.
Boundedness
The boundedness property is akin to stability and requires that in a linear problem
without sources the solution is bounded by the maximum and minimum boundary
values of the flow variables (Versteeg and Malalasekera 1996). Very simply, if the
boundary temperatures of a problem are 500 and 200 degrees then the interior values
of temperature should be less than 500 and greater than 200 degrees.
Transportiveness
Flow processes contain effects due to convection and diffusion. In diffusion
phenomena, such as heat conduction, a change in temperature at one location affects
the temperature in more or less equal measure in all directions around it. Convective
phenomena influences exclusively in the flow direction so that a point only
experiences effects due to changes at upstream locations. The transportiveness
property is a measure of the ability of the scheme to recognise the direction of flow.
The measure of these relative strengths are defined by the local cell Peclet number
(Versteeg and Malalasekera 1996). Differencing schemes dealing only with diffusion
will not require this property.
Accuracy
Classically the accuracy to which convection and diffusion terms are approximated is
judged via the truncation terms of the Taylor series expansion hence the order of the
scheme, although it should be noted that other features of the differencing schemes
make the assessment of accuracy less straight forward (Patankar 1980). For example
24
equation 3.8 is termed second order accurate as terms higher than second order are
truncated.
The following differencing schemes, which represent the most popular high and low
order schemes, will be discussed in terms of their suitability in dealing with the
convection terms.
In the control volume notation our attention is focused on a general node P with the
neighbouring nodes identified by W and E with the control volume faces denoted by w
and e. Further detail is given in Figure 11.
3.5.1 The Central Differencing Scheme
Many commercial CFD packages use this scheme for evaluation of the diffusive
fluxes. Generally this scheme is not appropriate for use in dealing with the convection
term in wind engineering simulations or indeed any high Reynolds number flow as it
is only stable and accurate for local cell Reynolds numbers of less than 2. For a given
viscosity the velocity must be very low or the grid size very small to achieve such a
Reynolds number. In addition, this scheme does not possess the transportiveness
property (Versteeg and Malalasekera 1996). The accuracy of the central differencing
scheme in terms of the Taylor series truncation error is second order. For further
information see Abbott and Basco (1989).
Referring to Figure 11, for a uniform grid we can write the west cell face property as:
2 / ) (
W P w
+ (3.11)
3.5.2 The Upwind Differencing Scheme
When the flow is in a positive direction, the west cell face is simply deduced from:
W w
(3.12)
This scheme satisfies the requirements of conservativeness, boundedness and
transportiveness. Accuracy is said to be first order in terms of the Taylor series
truncation error. One particular drawback with upwind differencing is apparent when
25
the flow direction is not aligned with the grid lines. This resulting error is often
termed false diffusion and is found to dominate the physical diffusion whenever the
Peclet number exceeds two, refer to Patankar (1980) for full details. It has been found
that when the leading term of the truncation error is an even order derivative the
numerical solution will appear to have an extra viscosity. A very simple explanation
of this effect can be given upon referring to the Navier-Stokes equations in section
3.2. These equations have terms such as

2
2
u
x
multiplied by the coefficient of
viscosity, , which represents the dissipative effect of the physical viscosity on the
flow. The even order derivatives in the finite difference formula also act as a
dissipative term much like the viscous terms in the Navier-Stokes equations. However
this is a consequence of numerical discretisation and as such is purely of numerical
origin (Anderson 1995). This artificial viscosity has the effect of increasing the
stability of a simulation by reducing high flow gradients due to the artificially high
viscosity generated. Unfortunately this effect also makes the scheme inaccurate.
3.5.3 Hybrid Differencing Scheme (Spalding 1972)
This differencing scheme is a combination of the central and upwind differencing
schemes reviewed earlier. Below local cell Reynolds numbers or Peclet numbers of 2
the central differencing scheme is used, above this value the upwind differencing
scheme is utilised. For computationally economical numbers of grid nodes used in
wind engineering applications the hybrid scheme defaults to the upwind scheme as the
Peclet number is greater than 2. The same accuracy limitations as previously
explained for the upwind scheme apply here although minor modifications and
improvements have been made, refer to Patankar (1980) for further details.
3.5.4 QUICK -Quadratic Upstream Interpolation for Convective Kinetics
(Leonard 1979)
In order to increase the accuracy of a differencing scheme it is necessary to involve
more neighbouring points and effectively bring in a wider influence of the
surrounding variables to deduce the next value of the dependent variable. In the
example given in section 3.4 the value of U at the point x was determined from two
neighbouring points. By increasing the number of neighbouring points it is possible to
reduce the discretisation errors. Linear interpolation between points is therefore no
26
longer a possibility. The next higher order interpolation, quadratic, is employed to
join the extra nodal points. Using the finite volume method the QUICK scheme
approximates the face value by quadratic interpolation between the immediately
adjacent nodal values and the next nearest upstream nodal value. This scheme has
many benefits as it satisfies the requirements of conservativeness and transportiveness
with third order accuracy in terms the Taylor series truncation error and it is not
affected by numerical viscosity (Lau 1987). However the QUICK scheme can cause
instabilities in a numerical simulation due to the fact that under and over shoots can
occur, as this scheme does not posses the boundedness property discussed earlier. In
complex turbulent flow fields such as those found in wind engineering the QUICK
scheme can sometimes lead to negative turbulence which is unphysical.
For a uniform grid the west cell face property is as follows:
WW W P w

8
1
8
6
8
3
+ (3.13)
3.5.5 CCCT - Curvature Compensated Convective Transport Scheme
(Gaskell and Lau 1988)
As a consequence of the quadratic profile the interpolation may overshoot the
physical boundary of the solution domain (see Figure 12). This problem led to the
development of the CCCT algorithm that maintains the beneficial features of the
QUICK scheme whilst providing boundedness and hence computational stability.
Rather than interpolating with a quadratic profile the scheme uses a monotonic profile
as shown in Figure 13.
The CCCT scheme has:
WW W P w

,
_

+
,
_

+ +
,
_


8
1
2
4
3
8
3
(3.14)
Where depends on the curvature of the variable . The scheme is then treated like
the QUICK scheme, refer to Gaskell and Lau (1988) for further details.
27
3.6 Calculation of the Flow Field and Pressure Correction
The solution of the momentum transport equations presents us with a number of
problems, further to those already highlighted, as follows:
1. The convection terms in the momentum equations that are derived from the
acceleration of a patch of fluid, contain non-linear quantities.
2. All three-momentum equations are intricately coupled because every velocity
component appears in each momentum equation and the continuity equation. The
most complex issue to resolve is the role played by the pressure. It appears in all
the momentum equations, but there is no equation for pressure.
Both these problems are associated with the non-linearities in the equation set. The
pressure velocity linkage can be resolved by adopting an iterative solution strategy
such as the SIMPLE algorithm (Patankar and Spalding 1972). In this algorithm the
convective fluxes per unit mass through the cell faces are evaluated from so called
guessed velocity components. Furthermore, a guessed pressure field is used to solve
the momentum equations and a pressure correction, deduced from the continuity
equation, is used to update the velocity and pressure fields. The iteration is started by
the use of an initial guess and proceeds until the accuracy of the initial guess is
improved sufficiently.
Full details of this method and the related problems which facilitate the need for
staggered grids and the Rhie and Chow (1983) algorithm can be found in Patankar
(1980).
3.6.1 Solving the Simultaneous Equations
In most cases, using commercial CFD codes, the discretised equations produced from
the partial differential equations are given in implicit form (for a full discussion on
implicit and explicit schemes see Shaw (1989)). When implicit schemes are used a set
of simultaneous equations are generated consisting of many individual equations. The
solution of these equations is very computationally intensive and as such a good deal
of work in the CFD community has focused on finding efficient solution techniques.
Techniques of solving the equations used in commercial CFD codes include Jacobi
and Gauss-Seidel methods including point relaxation and line relaxation techniques.
28
Advanced methods include Stones Strongly Implicit Procedure and matrix
manipulation method. All these schemes are detailed in full in Smith (1978), Hirsch
(1990), Abbott and Basco (1989) and Patankar (1980). One advanced method of
solving the simultaneous equations that is now very popular in the CFD community is
the multi-grid method. The philosophy behind this method is that the early iterations
are carried out on a fine grid and then progressively transferred to a series of coarser
grids. Since the coarser grids have less grid points fewer calculations are needed for a
given sweep (see section 3.6) of the flow field and the computational effort is
reduced. The results on the coarser grid are then transferred back to the fine grid and
the process is repeated a sufficient number of times until satisfactory convergence
(see section 3.6) on the fine grid is obtained (Anderson 1995). In mathematical terms
this method is able to more effectively damp the numerical errors as errors in a flow
field of wavelength approximately equal to the size of a cell are reduced much more
quickly. It is stated that by transferring the results from a fine grid to a series of
coarser grids the errors are more readily damped and faster solution convergence is
possible (Anderson 1995).
3.7 Summary
In this chapter the reader has been introduced to the basis of numerical fluid flow
calculations, the Navier-Stokes equations and the method of simplifying the effects of
turbulence resulting in the Reynolds stresses. Discussion has been presented on the
methods of discretising the partial differential equations, with special attention paid to
the treatment of the convection term. Finally, other important issues such as treatment
of the pressure term and solution procedures for the simultaneous equations have been
briefly discussed with suitable references offered for further reading.
29
4. Turbulence Modelling
4.1 Introduction
Engineering fluid mechanics is dominated by the fact that in most situations
of interest the fluid motion exhibits a chaotic non repeating unsteadiness
known as turbulent motion. At an engineering level the flows are too complex
for the details of turbulence to be resolved. Instead the equations of motion are
averaged over a time period producing what are known as Reynolds equations.
The approximate representation of these Reynolds stresses in the Reynolds
equations in terms of known or calculable quantities is known as turbulence
modelling (Launder 1989, p2).
The reader has previously been introduced to the Navier-Stokes equations (see section
3.2) and the necessity to make assumptions regarding these equations to allow for
calculations of turbulent flow. Nonetheless, the procedure of time averaging these
equations results in an additional set of terms, the Reynolds stresses, that have to be
accurately represented in some way. Whereas the original instantaneous Navier-
Stokes equations for laminar flow can be closed when the appropriate initial and
boundary conditions are prescribed, the time-averaged equations unfortunately
cannot. This is defined as the closure problem whereby further equations are required
to tie the Reynolds stress tensor to the mean flow equations. The level of closure
adopted refers to the number of supplementary transport equations required to achieve
closure of the Reynolds equations. This can include a number of turbulence modelling
practices that account for the convective and diffusive transport parameters including,
kinetic energy, turbulent vorticity and the Reynolds stresses themselves.
Accurate representation of the effects of the Reynolds stresses in the time averaged
Navier-Stokes equations is one of the most important aspects of a successful and
realistic CFD simulation. Therefore it is the aim of this chapter to cover the general
theory behind the most popular and widely available turbulence models, used at the
present time, and to undertake a critical analysis of each of the models using the
available published information.
30
4.2 Reynolds Averaged Navier-Stokes Equations (RANS)
The ensemble or time averaged form of the Navier-Stokes equations are called the
Reynolds Averaged Navier-Stokes equations (RANS). These models calculate a
mean, steady state velocity and pressure field and account for the velocity and
pressure fluctuations through additional modelled variables. These equations express
only the movement of large scale eddies thus allowing the use of coarse grids and
making the models relatively economical to use. A number of models are available
under this general heading which range from closure models based on the eddy
viscosity concept to full second moment closure models which represent the effect of
each component of the Reynolds stress tensor on the mean flow.
4.2.1 The Eddy Viscosity Concept
There are many theories regarding turbulence attributed to a number of researchers in
the 19
th
and 20
th
centuries. Early discoveries by Reynolds drew an analogy between
momentum transfer and the transfer of heat and matter for turbulent motion. Taylor
(1921) and (1938) provided the earliest theoretical and mathematical explanation for
the highly diffusive nature of turbulence. The pioneering work of Kolmogorov (1941)
led to a greater understanding of how turbulence interacts with the mean flow, the
energy cascade and its dissipation. In terms of the turbulence modelling perhaps the
most important research is attributed to the earliest worker in this field, Boussinesq
(1877), who postulated that the Reynolds stresses should be proportional to the mean
strain rate. This concept is based on the assumption that both the viscous stresses and
the Reynolds stresses act on the mean flow in a similar manner. Referring to Equation
3.5, inserted below for clarity, it can be seen that both these stresses appear on the
right hand side of the momentum equation.
i j i
i
j
j
i
j i j
i j
i
F u u
u
x
u
x x
u u
t
u
+
x
+
x
+
) (
=
) (
1
1
]
1

,
_


Furthermore:
In Newtons Law of Viscosity the viscous stresses are taken to be
proportional to the rate of deformation of fluid elements (Versteeg and
Malalasekera 1996, p63)
31
For an incompressible fluid this relationship can be expressed in mathematical terms
as:

ij ij
i
j
j
i
e
u
x
u
x
= = +

_
,

(4.1)
(Versteeg and Malalasekera 1996)
Also
It is experimentally observed that turbulence decays unless there is a shear in
isothermal incompressible flows. Furthermore turbulent stresses are found to
increase as the mean rate of deformation increases. (Versteeg and
Malalasekera 1996, p63)
Therefore, these statements led Boussinesq to propose a linear relationship between
Reynolds stresses and the rate of deformation of a fluid linked by a coefficient of
proportionality
t
as follows:

,
_


i
j
j
i
t j i ij
x
u
x
u
u u

+ = = (4.2)
(Versteeg and Malalasekera 1996)
Boussinesqs reasoning seems to be logical when one considers that the energy
dissipation and transport of mass momentum normal to the flow in laminar flows are
all mediated by viscosity. As the effect of turbulence is to greatly increase this process
it seems a natural assumption to conclude that an extra viscosity can adequately
represent the effects of turbulence (Versteeg and Malalasekera 1996).
The right hand sides of Equations 4.1 and 4.2 are the same except for the coefficient
used linking the two sides of the equations. The main difference between these two
coefficients is that is a function of the fluid properties only while
t
is a function of
the turbulence.
If Equation 4.2 is substituted into Equation 3.1 then the mean flow equation now has
an enhanced additional viscosity
t
due to the turbulence of the flow. Using this
approach the modelling process can be completed if the turbulent viscosity can be
found from other variables. For a full derivation of the eddy viscosity concept see
32
Gatski et al (1996). There are a number of methods available of deriving the value of
turbulent viscosity as will be briefly detailed in the following sections.
4.2.2 The Mixing Length Model
In the mixing length model it is argued that at a given point and time in the flow field
turbulence can be characterised by a single representative velocity scale and length
scale. Hence the turbulent viscosity is expressed in terms of these two scales as a
function of position as follows:

,
_

x
+ =
2



v
y
u
l C
t
(4.3)
(Launder 1972)
The mixing length model is of use if the convection and diffusion of turbulent
properties can be neglected (Versteeg and Malalasekera 1996). In fact these properties
are very important for wind engineering flow fields, which include recirculatory
flows. Consequently more complex statements are required involving fluid transport
equations which may express these effects in terms of the dynamics of the turbulence.
Full details of the mixing length model can be found in Lectures in Mathematical
Turbulence by Launder (1972).
4.2.3 The Standard k- Turbulence Model
The standard k- model (Launder and Spalding 1974) has two model transport
equations, one for the turbulent kinetic energy of the flow, k and one for the
dissipation rate of k, . These values are used to define the velocity scale and the
length scale, at any given point and time in the flow field, representative of large scale
turbulence as follows:
Velocity scale = k
1
2
(4.4)
Length scale l
k
=
3
2

(4.5)
where k = turbulent kinetic energy
= the dissipation of turbulent kinetic energy
33
From this the eddy viscosity can be specified as follows:



2
= . . =
k
C l C
t
(4.6)
Inserting the Boussinesq hypothesis into the momentum equation yields

x
u
+
x
=
1
1
]
1

,
_

+
i
j
j
i
eff
j i j
i j
i
x
u
x
P
x
u u
t
u

(4.7)
where;

eff
= +
t
. (4.8)
The standard k model equation is obtained by multiplication of the instantaneous
Navier-Stokes equations by the appropriate fluctuating velocity components (i.e. x-
component equation multiplied by u etc.) and addition of all the results. This is
followed by a repeat of this process on the time averaged Reynolds equations,
subtraction of the two resulting equations and substantial re-arrangement yielding the
equation for the turbulent kinetic energy k (Tennekes and Lumley 1972). It is also
possible to develop similar transport equations, from the Navier-Stokes equations, for
other turbulence quantities including the rate of viscous dissipation . Nonetheless it
should be noted that the energy dissipation equation is far more empirical and the
modelling of terms is so severe that it is best to regard the entire equation as a model.
The standard k- model equations are (including Equation 4.7) as follows:
Turbulent kinetic energy
( )
1
1
]
1

j j j
i
ij
j
j
x
k
x x
u
x
k
u
t
k

k t
+ + = + (4.9)
Dissipation rate
( )
1
1
]
1

j i j
i
ij
j
j
x x k
C
x
u
k
C
x
u
t

t
2
2 1
+ + = +
(4.10)
34
where
ij ij ij
k S
3
2
2 =
t
= Reynolds stress tensor.
and
ij
= the Kronecker delta (1 when i = j)

,
_

i
j
j
i
ij
x
u
x
u
S
The various constants in the above equations are necessary due to the numerous
simplifications made to the models and are derived from comprehensive data fitting to
a wide range of turbulent flow fields (usually wind tunnel data).
To calculate the Reynolds stress tensor in the k- model a revised Boussinesq
relationship is used from that shown in Equation 4.2.

3
2
+ =
ij
i
j
j
i
t j i
k
x
u
x
u
u u

,
_


(4.11)
(1) (2) (3)
The effect of this extra term (3) added to the Boussinesq relationship is to make the
term applicable to the normal Reynolds stresses as the standard hypothesis deals only
with shear stresses. This term effectively allocates an equal third of the sum of the
normal Reynolds stresses to each normal stress.
The result of the assumptions and simplifications discussed in this section, which
result in a isotropic, scalar, description of the eddy viscosity coefficient,
t
, are
discussed below.
4.2.3.1 Discussion
Of all the available turbulence models, the k- model is by far the most widely used
and has been tested for a vast number of flow fields. It is favoured in industrial
applications due to its relatively low computational costs and generally better
numerical stability than more complex turbulence models such as the differential
stress model (see section 4.2.7). This model has proved a success in many
35
applications, particularly in confined flows where the normal Reynolds stresses are
relatively unimportant. Unfortunately the opposite is true of wind engineering flow
fields and the k- model performs poorly.

It is explained in a number of research papers (Murakami 1997, Murakami et al 1991)
that the k- model incorrectly estimates the level of turbulent kinetic energy around a
surface mounted cube, in particular at the leading top edge (see chapter 6). This
fundamental error results in a poor representation of the flow field and pressure
distribution around the cube. In addition, wind tunnel tests have deduced that the
streamwise normal stress u u dominates in the area of the roof and in the downstream
free shear layer. Also the lateral normal stress v v dominates u u , with w w the
smallest, in the wake recirculation zone, where a Karman vortex street occurs
(Murakami 1990). Consequently the turbulent flow field in wind engineering
applications is most definitely anisotropic. Therefore the assumption of a simple
isotropic eddy viscosity term is insufficient to adequately describe the complexity of a
highly anisotropic flow field and results in the k- models failure to accurately predict
many turbulent flow fields, not least in wind engineering applications. The main
source of error in this model is therefore rooted in the linear Boussinesq
approximation and the isotropic eddy viscosity concept.
Full testing of this model and comparison with full-scale data is available in chapter 6.
4.2.4 The Low Reynolds Number k- Model
The equations for this model are only slightly modified from the standard k- model,
the main difference is apparent in the treatment of the near wall region. The high
Reynolds number k- model reduces the computational effort of a given flow
simulation by making use of the universal behaviour of near wall flows (see section
4.3). The low Reynolds number model does not use this method and effectively
integrates to the wall surface in the low Reynolds number region of the flow. The
standard model therefore has to be revised to effectively force the correct near wall
conditions whereby the viscous stresses in the near wall region take over from the
Reynolds stresses which are dominant in the flow at a much greater distance from the
wall. This is achieved by the use of wall damping functions that multiply the model
constants C C

,
1
etc.
36
For example:



2
=
k
f C
t
(4.12)
where

,
_

2
)
50
1 (
4 . 3
exp
T
R
f

(4.13)
and

2
k
R
T
(4.14)
Further damping functions are applied to the turbulence transport equations. Similar
functions can be applied to any of the other models detailed in this section to allow
them to resolve the flow in the low Reynolds, near wall region, of the flow.
For further details see Patel et al (1985).
4.2.5 The k - Equation Model
This turbulence model, first proposed by Kolmogorov (1941), was in fact the first
model of turbulence. The variable k is, as usual, the turbulent kinetic energy and is
the dissipation per unit turbulent kinetic energy. In the usual manner these two terms
are modelled using partial differential equations. The advantage of replacing the
equation with the equation is that the second is easier to integrate (more robust) and
that it can be integrated through the sub-layer without the need for additional damping
functions (Menter and Grotjans 1999)
Eddy viscosity


k
t
= (4.15)
37
Turbulent kinetic energy
( )
1
1
]
1

j
T
j j
i
ij
j
j
x
k
x
k
x
u
x
k
u
t
k

* *
+ + = + (4.16)
Specific dissipation rate
( )
1
1
]
1

j i j
i
ij
j
j
x x x
u
k x
u
t

T
2
+ + = + (4.17)
(Wilcox 1993)
where
Although experimental evidence of the performance of this model is relatively scarce,
Wilcox (1994) showed that for a two dimensional backward facing step the
reattachment length of the recirculation zone was within three percent of the
experimentally measured location. This out performed the standard k- model, which
significantly underestimated the reattachment length. Disadvantages with this model
include the fact that the solutions produced are very sensitive to the values specified
for at the inlet. Menter has proposed a model that combines the advantages of the k-
and k- models, thereby removing this deficiency (CFX-International 1998).
At present there is little information available on this model and its application to
wind engineering flow fields, consequently the author has included this model in tests
in chapter 7.
4.2.6 The Renormalisation Group (RNG) k - Turbulence Model
An alternative way to derive turbulence closure models was proposed by Yakhot and
Orzag (1986). They applied Renormalisation Group theory to the Navier-Stokes
equations and derived a two equation k- model. Further details of the RNG
technique can be found in McComb (1990).
ts coefficien closure are etc. , ,
*

38
Referring to Equation 4.10, detailing the equation for the standard k- model, it can
be seen that the only revision to this equation for the RNG model is the inclusion of
the extra term C
1RNG
and the inclusion of revised model constants:
( )
1
1
]
1


j i j
i
ij RNG
j
j
x x k
C
x
u
k
C C
x
u
t

t
2
2 1 1
+ + ) ( = +
(4.18)
where
C
RNG 1
0
3
1
1

( )
( )
(4.19)
and

_
,

P k
T
1
2
(4.20)

0
and are additional model constants.
The remaining equations for this model are identical to the standard k- model.
4.2.6.1 Discussion
Unfortunately it appears that there is scant information available regarding testing of
this model in wind engineering flow fields. Therefore a full test has been conducted
by the author to assess the performance of this model, see chapter 6 for further details.
Furthermore, although this model has been given impressive reviews by many
researchers (refer to Orzag 1994) it is important to remember that it is only a variant
of the k- model and as such is still based on the flawed isotropic eddy viscosity
assumption and Boussinesq approximation. In a similar manner to many other models
that are modified versions of the standard k- model it may provide improved
predictions for some applications while giving lower accuracy for others.
This basic feature of these models led to a review of a turbulence model that does not
employ the eddy viscosity concept to describe the effects of the Reynolds stresses but
provides expressions for all the individual components of the Reynolds stress tensor.
39
4.2.7 The Differential Stress Equation Model (DSM)
A more complex version of the RANS equations is the differential stress model of
Launder, Reece and Rodi (1975).
Second moment closure is based on exact transport equations for the individual
Reynolds stresses and fluxes derived from the Navier-Stokes equations. The
derivation of the Reynolds stress equations, detailed in full by Leschziner (1990),
results in a total of nine transport equations, six of which describe the Reynolds
normal stresses and shear stresses and a further three which describe the Reynolds
fluxes, which for an incompressible fluid gives:
( ) ( )
- - - - - - - - - - - - d - - - - - - - - - - - - - - - - - -
2
- - - - - - F - - - - - - - - P - - - - - - - - C - - -

ij ij
ij ij ij ij

,
_

+ +

,
_

+ + + +

,
_

+
k
j i
ik j jk i k j i
k k
j
k
i
i
j
j
i
i i i i
k
j
k i
k
i
k j j i k
k
x
u u
pu pu u u u
x x
u
x
u
x
u
x
u
p f u f u
x
u
u u
x
u
u u u u u
x
(4.21)
This translates to
F + P =
ij ij ij ij
C +
Transport of Reynolds Production of R.S Stress production Redistribution
stresses by convection or distruction by of R.S due to
action of rotational pressure - strain
body forces interaction
ij ij
d
Rate of dissipation Rate of diffusion.
by viscosity
Wilcox (1994) details the individual contents of the equations and lists some of the
reasons for improvements over the eddy viscosity models as follows.
40
Firstly, since the equation automatically accounts for the convection and diffusion of
the Reynolds stresses, a second order closure model will include effects of flow
history. In turbulent shear flow large bodies of fluid migrate across the flow, carrying
smaller scale disturbances. In addition to migrating across the flow, they have a
lifetime so long that they persist for distances as much as 30 times the width of the
flow (Bradshaw 1973). Thus the turbulent stresses at a given point depend upon
upstream history and cannot be uniquely specified in terms of the local strain rate
tensor, as is the case with the eddy viscosity formulation, and the linear Boussinesq
relationship.
Secondly, Equation 4.21, being more complex than the eddy viscosity approximation,
contains a greater number of terms enabling a greater number of flow effects to be
accounted for. The extra terms such as convection, production and body force allow
representation of flows involving streamline curvature, system rotation and
stratification. Thirdly, there is no reason why this model should give equal values for
the normal and shear stresses as the individual components are calculated separately.
In a similar fashion to the models described previously, there is now a significant
closure problem with the Reynolds stress model. In order to close the Reynolds stress
equations it is necessary to eliminate the third moment correlations that appear in
Equation 4.21 of the form u u u
k i j
. Furthermore, it is necessary to model the
dissipation tensor
ij
, the turbulent transport tensor C
ij
and the pressure strain
correlation tensor
ij
in the Reynolds stress transport equation. As each of these
forms is a tensor the approximation required for closure can assume much more
elaborate forms compared to approximations used in the k equation (Wilcox 1994).
The diffusion term in the Reynolds stress transport equation is often modelled using
Kolmogorovs (1941) hypothesis of local isotropy of small scales where directionality
of the small scale eddies is damped due to the effects of viscosity. The model is
formulated so that the normal Reynolds stresses only are affected (Versteeg and
Malalasekera 1996).

ij ij
=
2
3
( i j ) (4.22)
where

,
_

k
i
k
i
x
u
x
u
v

= dissipation rate of the turbulent kinetic energy


41
It should be noted that this assumption is not always true and the dissipation of
turbulent kinetic energy may, in certain circumstances, be anisotropic. Rotta (1972)
has attempted to account for this.
The turbulent transport term is often modelled by the analogy that the rate of transport
of Reynolds stresses by diffusion is proportional to the gradients of the Reynolds
stresses.
The pressure-strain correlation in the differential stress model has proven the most
difficult term to account for and has received the greatest amount of attention from
turbulence modellers. The physical processes that cause pressure fluctuations to occur
are detailed and complex and include interaction between neighbouring eddies and the
effects of different mean velocities on the turbulent eddies (Tennekes and Lumley
1972). The main effect of this term is to reduce the level of the Reynolds shear
stresses and redistribute this energy amongst the normal stresses. This returns the
normal stresses towards a state of isotropy. Unfortunately the presence of a solid
boundary tends to increase the anisotropy of the turbulent eddies and decrease the
magnitude of the Reynolds shear stresses in contrast to the role of the pressure strain
term. Modifications to the pressure-strain term are thus required to enable the model
to accurately predict the effect of the wall on the turbulent eddies. Further details are
given in chapter 6.
4.2.7.1 Discussion
The differential stress model has a far greater universality than the models based on
the eddy viscosity concept due to its more rigorous and detailed mathematical
formulation. The inclusion of a greater number of equations allows for a far greater
description of the physics of turbulent flow.
Despite the considerable abilities of this model it still has many inadequacies and
there are many years of development work to be undertaken on the differential stress
model and its closure forms before it will show its real potential. At present the main
areas identified as causing inaccuracies in flow predictions include the following:
1. the modelled turbulence energy dissipation equation

42
2. the closure form of the pressure-strain tensor and the effect of wall reflection terms

3. numerical instability and difficulties in obtaining converged solutions for the
modelled partial differential equations in complex flows.
The turbulence energy dissipation equation used in the differential stress model is the
same as that used in the k- model and is said to be a source of considerable error.
There is presently a lack of understanding regarding the whole energy spectrum for
the dissipative process. The exact energy dissipation equation is significantly more
complicated than the equation used in the model and involves terms such as
production of dissipation, dissipation of dissipation, the sum of molecular diffusion of
diffusion and turbulent transport of dissipation (Wilcox 1994). It is explained that
these correlations are essentially impossible to measure using conventional
experimental techniques and that only direct numerical solution techniques will shed
light on these terms.
Point number two is detailed further in section 5.5.
The third issue is particularly important with regards to general application of this
model. As the time scales associated with the turbulence are much shorter than those
connected with the mean flow, the equations for both the k- model and DSM are
much stiffer than the laminar flow equations, therefore the solution method has to take
account of this. An equation is said to be numerically stiff when there are two or more
very different scales of the independent variable on which the dependant variables are
changing. For this reason, in the numerical solution procedure, one first performs an
outer iteration of the momentum and pressure correction equations in which the value
of the eddy viscosity or Reynolds stresses are based on the solution to the turbulence
transport equations from the preceding iteration. Then an inner iteration of the k, or
Reynolds stress equations is performed following linearisation of the model
equations. The outer iteration with an updated stress or eddy viscosity is then repeated
and so on. The stiffness of the equations is the reason why they are treated separately
as described above, coupling the equations would make convergence very difficult to
obtain. Too large a time step (or its equivalent in an iterative method) can lead to
negative or unrealisable values for the turbulence model equations and thus
instability. The number of equations and time scales involved in describing turbulence
with the DSM has led to even greater numerical stiffness and the need for great care
with this model in order to obtain accurate solutions.
43
4.2.8 The Algebraic Stress Model (ASM)
The algebraic stress model is an economical way of accounting for the anisotropy of
the Reynolds stresses without going to the full length of solving the Reynolds stress
transport equations. Removing or modelling the convective and diffusive transport
terms effectively reduces the Reynolds stress terms from transport equations to a set
of algebraic equations. The general method chosen to model these terms is to assume
that the sum of the convection and diffusion terms of the Reynolds stresses is
proportional to the sum of the convection and diffusion terms of the turbulent kinetic
energy. The resulting algebraic approximation of the Reynolds stress takes the
following form:
( )

k
P P
P
C
C
k u u
ij ij
D
ij j i

3
2
1
3
2
1
' '

,
_

+
+ (4.23)
where C
D
is an adjustable constant dependant on the type of flow field encountered
and C
1
= 2.2.
4.2.8.1 Discussion
Although this model is reported to allow for anisotropic predictions of turbulence and
is only marginally more computationally expensive than the standard k- model it
does have significant problems which have led the author not to include the model in
the tests in chapter 6. The ASM is subject to all the disadvantages as described for the
DSM, including the numerical stability problems, and it is reported to be severely
restricted in flows where the transport assumptions for the convective and diffusive
effects do not apply. Murakami et al (1991) has reported that the modelling
assumptions for these terms produce particular inaccuracies in wind engineering flows
where the convection and diffusion of the flow are extremely important. Furthermore
the model has to be tuned to different flow fields through the adjustable constant
detailed above. For further information the interested reader is referred to Murakami
(1993).
44
4.3 Turbulent Wall Boundary Conditions
As briefly explained in section 4.2.4, a fundamental part of a CFD simulation is the
method used to represent the effects of the wall on a turbulent flow. The majority of
the turbulence models discussed will be used in high Reynolds number flows. This
requires an economical method of modelling the effects of the wall on the momentum
and turbulence transport equations. This is particularly important in wind engineering
applications as these are often high Reynolds number turbulent flows with complex
wall bounded geometries.
Normally wall boundary conditions are specified using wall functions. This is
necessary to avoid the need for very fine grids to resolve the large gradients of energy
dissipation in the near wall region and thus reduce the computational overheads of a
given wall bounded problem. Wall functions are based on the universal assumptions
that a constant shear stress exists in the near wall region and that the length scale of a
typical eddy in this region is proportional to the distance from the wall. These
assumptions result in a logarithmic velocity profile near the wall
The wall law relates the shear stress to the turbulent kinetic energy:
k C . .
2
1

(4.24)
This is used to define a velocity scale:
( )
u u .
.
2
1

+
(4.25)
and the scaled wall distance:

p
y
y

+
(4.26)
For any high Reynolds number turbulence model the implementation of wall
boundary conditions starts with the evaluation of y
+
(from Equation 4.26) where
p
y
is the distance of the near wall node to the solid surface. A near wall flow is taken to
be laminar if 63 . 11
+
y and the wall shear stress is assumed to be entirely viscous in
origin. If 63 . 11 >
+
y the flow is turbulent and the wall function approach is used and
the finite volume node is considered to be in the log law region of the turbulent
45
boundary layer. In this region wall function formulae associated with the log law are
used to calculate shear stress, heat flux and other variables. High Reynolds number
turbulence models usually employ the assumption that if y is the co-ordinate direction
normal to the wall, the mean velocity at a point y
p
with 30<y
+
<500 satisfies the log-
law. Measurements of turbulent kinetic energy budgets indicate that the rate of
turbulent production equals the rate of dissipation. Using these assumptions and the
eddy viscosity formula it is possible to develop the following wall functions (Versteeg
and Malalasekera 1996):
( )

'

>
<

+ + +
+ + +
+
0
0
, . log
1
,
y y for y E
y y for y
u

(4.27)

2

C
u
k

(4.28)
y
u

(4.29)
where
+
0
y defines the cross over point between the laminar sub-layer and the
logarithmic region. E is the log layer constant and is the Karman constant, both of
which are empirical values found from experiments. If the walls were not smooth, as
discussed above, E should be adjusted accordingly and a new limiting value of y
+
would result. See section 6.3.7 for details.
The wall function approach is not completely satisfactory for several reasons. Most
importantly, numerical solutions are sensitive to the point above the surface where the
wall functions are used, i.e. the point where the matching occurs. In addition, the law
of the wall does not always hold for the flow near solid boundaries, most notably for
separated flows (Wilcox 1994). Further discussion of this point is given in chapter 7.
More details on the individual layers of the near wall region are given below.
4.3.1 Linear (or laminar) sub-layer the fluid layer in contact with the wall
At the solid surface the fluid is stationary and turbulent eddying motions also stop
very close to the wall. In the absence of turbulent (Reynolds) shear stress effects the
46
fluid closest to the wall is dominated by viscous shear. This layer is in practice
extremely thin (
+
y <5) and we may assume that the shear stress is approximately
constant and equal to the wall shear stress
w
throughout the layer. As there is a
linear relationship between
+
u and
+
y this region is referred to as the linear sub-
layer.
4.3.2 Log-law layer The turbulent region close to a smooth wall
Outside the laminar sub-layer a region exists (30 ) 500 < <
+
y where viscous and
turbulent effects are both important. The shear stress varies slowly with distance from
the wall and within this region it is assumed to be constant and equal to the wall shear
stress.
) ln(
1
ln
1
+ + +
+ Ey B y u

(4.30)
Numerical values for the constants are found from experiments. Wall roughness
causes a decrease in the value of B as highlighted in section 6.3.7. As there is a
logarithmic relationship between
+
u and
+
y this region is referred to as the log-law
layer.
4.4 Alternatives to the Navier Stokes Equations
4.4.1 Chaos
The mathematical theory of Chaos has been the subject of much research work in
recent years. Wilcox (1994) states that mathematicians have discovered that certain
non-linear dynamical systems, sets of non-linear ordinary differential equations,
possess an extremely complicated chaotic structure. Researchers in this field have
successfully used these systems to describe very simple turbulent fluid flow processes
such as convection between two horizontal plates of different temperatures, known as
Rayleigh-Benard convection. For detailed theories on this subject see Gleick (1988),
Deissler (1989) and Stewart (1989).
There are many problems in using this very complex and theoretical approach to
model turbulence. Bradshaw (1992) has shown that solving turbulent flow problems
47
using these methods would require an amount of computing time equivalent to that
required for LES. Furthermore, it has been deduced there is more to turbulence than
randomness or chaos. As is explained later (see section 5.7.2) so called coherent
structures exist to some degree in a majority of turbulent flows. If turbulence was
simply random then perhaps statistical methods would be able solve the problem.
4.4.2 The Discrete Vortex method
This technique is a much more viable alternative to the Navier-Stokes equations than
Chaos theory. Very generally, the method is to assume a form of vortex, for example
a point vortex, and then to consider an assembly of a few hundred or thousand such
vortices. In this way turbulent shear layers can be reproduced. However, there are
problems in adequately representing boundary conditions in complex flows (Quinn
1999) as well as the fact that flows are represented as being purely invisid, which is
generally not appropriate for widespread CWE application.
For further details see McComb (1990), Kawai (1990) and Leonard (1985).
4.5 Summary
This chapter has introduced the reader to the concept of turbulence modelling and the
various methods used to represent the effects of the turbulence on the mean flow or
Reynolds averaged Navier-Stokes equations. A number of schemes have been
reviewed ranging from simple eddy viscosity approximations, such as the mixing
length model, through to more complex models such as the differential stress model.
All of the techniques discussed have been subjected to a critical appraisal to allow a
focused discussion, in the following chapter, of a number of new turbulence models
that may offer possible solutions or improvements to the main problems areas.
48
5. Developments in Turbulence Modelling.
5.1 Introduction
In the review of turbulence models in chapter 4 it was found that the main problem
areas were as follows:
1. the isotropic eddy viscosity assumption and linear Boussinesq hypothesis used in
the standard two equation turbulence models
2. numerical instability of the DSM and errors in the pressure-strain term
3. errors in the dissipation equation.
Unfortunately at the present time there are few, if any, improvements available for
point 3 that can be applied to the dissipation equation and used in a generic manner.
In addition, there is generally little that can be done to improve the numerical
instabilities in the DSM. As this project uses a commercially available CFD code it is
apparent that the developers have attempted to make the use of this package as user
friendly as possible and have in some cases favoured numerical stability over
accuracy (Wilkes 1999).
Consequently, excluding points 2 and 3, this chapter will expand upon the modelling
errors discussed in chapter 4 and highlight improvements and developments in
turbulence models to account for these defects. Wherever possible a critical appraisal
will be given. Further model improvements beyond those highlighted already will also
be discussed.
5.2 The Murakami, Mochida, Kondo (MMK) k- Turbulence Model
(Tsuchiya et al 1996)
Many techniques have been applied by scientists and researchers to improve the
results obtained by the standard k- model due to the deficiencies of the eddy
viscosity concept. The technique which is by far the most popular is to make flow
specific ad-hoc modifications to the model closure constants to force it into agreement
with experimentally derived flow fields. This technique, although popular, has the
49
effect of reducing the universality of the model. There are many published revisions
and ad-hoc changes to the model that successfully represent improvements over the
standard model for certain flow fields. The following section will describe one of the
more successful and extensive revisions appropriate to wind engineering flow fields,
the MMK k- model of turbulence.
The paper detailing this model proposes a k- model that is said to resolve the
problems encountered with the standard model (explained in section 4.2.3) by
modifying the expression for the eddy viscosity approximation. The main k and
model equations are unchanged. The revised k- model is then applied to flow fields
around bluff bodies including a surface mounted cube both normal and skewed to the
flow.
5.2.1 Outline of the revised model
where .
2
2


k
C s v P
t t k
(5.1)
The turbulent kinetic energy production term and the calculation of the eddy viscosity
for the standard k- are listed in Equation 5.1
For the MMK model, P
k
and
t
are as per Equation 5.1 except that:
1 of for values <

s s
C C

(5.2)
. 1 for

s
C C

(5.3)
(Tsuchiya et al 1996)
In the standard k- model, k is overestimated near the front edge of a surface mounted
cube which gives rise to a large eddy viscosity. This discrepancy is caused by
overestimation of the turbulence production term P
k
which is caused by the eddy
viscosity concept as described by Murakami et al (1992) as follows.
50
The full term (2-dimensional) for P
k
:
x
w
u w
z
u
w u
z
w
w w
x
u
u u P
k

(5.4)
This term equates P
k
to the difference between the turbulence production from
diagonal elements of the strain rate tensor and production from off diagonal elements.
The production term for the normal component of the Reynolds stress is as follows:
x
u
w w u u P
kn

) ( (5.5)
Using the continuity equation and then the eddy viscosity concept this equation
becomes:
P v
u
x
kn t

_
,
4
2

(5.6)
When an anisotropic model is used such as the differential stress model, P
kn
is
calculated using the form as described in Equation 5.5 which involves two velocity
components. When using the eddy viscosity concept these two normal stresses cannot
be incorporated so they are simply expressed as Equation 5.6 by adding the two
components of turbulence production (Murakami et al 1992). Hence the value of
turbulent production is always large and positive. Although the MMK model is still
constrained by the eddy viscosity concept it attempts an ad-hoc correction to the
over production of P
kn
. The standard model constant C

becomes a variable that can
reduce in magnitude dependent on the ratio of vorticity to shear. At flow impingement
areas where there are high shear stresses the ratio is less than one and a reduced value
of C

is calculated thus reducing the eddy viscosity returned.
Tsuchiya et al (1996) conducted tests using the MMK and the standard k- model for
flows over a two dimensional square rib and a three dimensional cube. It has been
reported that the MMK model out performs the standard model for wind engineering
flow fields in all aspects, including a better distribution of surface pressures.
Although this model has had encouraging reports it should be borne in mind that it is
still based on the fundamentally flawed assumption of isotropic eddy viscosity as used
by the linear, standard, k- model. This model could be described as an ad-hoc
51
modification to force the standard k- model into agreement with wind engineering
flows as the additions to the equations are not derived in any way from the Navier-
Stokes equations.
Although it is clear that this model has been previously tested in wind engineering
flow fields the author has conducted full tests in chapter 6 to obtain a full appreciation
of this technique.
5.3 Two Layer Turbulence Models
Low Reynolds number turbulence models, so called because they are able to integrate
into the near wall low Reynolds number region of the flow, are reported to require
high degrees of numerical resolution in the sub-layer. This is due to the need to
account for the steep gradients of the energy dissipation term, (Patel et al 1985). As
such the computational requirements of this model in complex three-dimensional wall
bounded flows can become excessive, as previously discussed.
Due to these constraints, and in an attempt to improve the performance of turbulence
models by taking a greater account of the effects of the viscous sub-layer, two layer
turbulence models have been formulated. These models make use of either a standard
k- or full Reynolds stress model to describe the flow field in the outer region and a
simpler model to describe the near wall region. The advantages of this method are
twofold. Firstly a more accurate description of the sub-layer is given than with the use
of wall functions and secondly a lower degree of numerical resolution is needed than
with the low Reynolds model. In order to avoid the resolution problems encountered
with the low Reynolds number model the sub-layer turbulence model usually avoids
the use of the transport equation to determine in favour of determination from a
prescribed length scale.
The sub-layer models tested and documented at this time consist of either a mixing
length model, which is applicable for both the inner and outer flow regions, or more
usually a one-equation model (Malecki et al 1993). For the one equation model a
value of the eddy viscosity in the near wall region is calculated from:
v C k l
t


1
2
(5.7)
52
Only one partial differential equation is solved in this model as the dissipation rate,
is determined from a prescribed length scale distribution:

k
l
1
2
(5.8)
The length scales l

and l

are described by exponential functions that express a


linear variation in the log law region until very near the wall where deviations occur
(Rodi 1990). The length scale l

reduces rapidly in the region very close to the wall


thus reducing the value of eddy viscosity determined in Equation 5.7.
The use of two layer turbulence models with the one equation model as described
means that there are now two descriptions of the turbulent kinetic energy k.
Consequently the natural matching point of the two models appears to be at the outer
edge of the log law sub-layer. Exact methods of matching the two models range from
the specification of exact grid points away from the wall to the fulfillment of certain
specific criteria. These can include such factors as ratios of eddy viscosity to
molecular viscosity and the value of l

, damping function relation close to unity, i.e.


very small viscous effects (Rodi 1990).
5.3.1 Discussion.
Experimental evidence suggests that the sub-layer velocity distribution deviates from
the logarithmic assumption for certain flow fields such as those including strong
secondary flows and separated regions, which may extend into the sub-layer (Rodi
1990). Therefore it could be argued that a more rigorous determination of the sub-
layer is required than that provided by simple wall functions.
Rodi (1990) has undertaken testing of k- and Reynolds stress two layer models for a
number of flow problems including those of interest in the field of wind engineering.
The flow over a 2D backward facing step was analysed with the aid of a two layer k-
model and compared with the standard k- model and experimental results. The
standard model was found to drastically under predict the reattachment length of the
separation and recirculation zone. The two-layer model was found to improve the
predictions with better predictions of velocity and shear stress profiles and
reattachment length. Although it was noted that the results where still not particularly
accurate with a 14 percent under prediction of the reattachment length for an inclined
53
upper wall case. Full details of tests with this model, in the field of wind engineering,
can be found on reference to Lakehal and Rodi (1997).
This model is again a simple isotropic eddy viscosity model with more attention paid
to the sub-layer. Furthermore, wall functions are used to calculate the energy
dissipation term so it could well be argued that this model is not truly integrating to
the wall as the low Reynolds number models.
Finally the author has not tested these models any further due to their high
computational overheads (detailed in chapter 6) and availability of previously
published results.
5.4 Revisions to the Boussinesq Hypothesis The Non-Linear k- Model
An effective viscosity hypothesis relates the Reynolds stresses solely to the rates of
strain of the fluid and to scalar quantities. As described in section 4.2.1 the first
effective viscosity hypothesis was proposed by Boussinesq as early as 1877. This
formula, which simply represents the action of the uw shear stress, has been used
with considerable success by, among others Ng (1971) and Rodi (1972) for free shear
flows. Nonetheless it has been observed that the Boussinesq hypothesis fails in a
number of applications including boundary layers over curved surfaces. Bradshaw
(1973, 1992) has stated that this failure is due to the form of the stress strain relation
rather than the inapplicability of the eddy viscosity approach.
In order to account for flows in which more than one Reynolds stress is required to
fully describe a given flow field the Boussinesq hypothesis is generalised to give the
isotropic eddy viscosity assumption used in the majority of turbulence models
(described in section 4.2.1). It has been shown in previous sections that this theory is
invalid for flows in which ' ' ' ' ' ' w w v v u u . The first attempts to remove this
deficiency in the eddy viscosity assumption were undertaken by Lumley (1970). Pope
(1975) adopted a similar approach to Lumley in formulating a constitutive relation for
the Reynolds stresses resulting in a finite tensor polynomial to form a revised general
effective viscosity hypothesis.
In deriving a new relationship Lumley (1970) shows that the mean velocity field and
boundary values of the fluctuating velocity are sufficient to determine the Reynolds
stresses and assumes that, far from boundaries, the boundary conditions serve at most
54
to set the levels of the time and length scales. With this assumption it follows that in a
homogeneous or near homogenous flow, where the rates of strain contain all the
information about the velocity field, the Reynolds stresses are a function of the rates
of strain and scaling parameters only (Pope 1975). Furthermore, these two scaling
parameters are sufficient provided that all the macroscales of turbulence are
proportional i.e. at high Reynolds numbers when the influence of laminar viscosity
may be excluded. Therefore using these arguments, applying dimensional analysis
and imposing invariance under coordinate transformation Pope (1975) suggested the
following form consisting of up to a fifth order relationship:


10
1
' '
3
2

ij ij ij
j i
ij
T G
k
u u
a (5.9)
where
[ ] ) , ( of Invariants , ,
) , (
ij ij ij ij
ij ij ij ij
S k G G
S T T

Pope (1975) states that the non-linear eddy viscosity has the following advantages
over the algebraic stress model, with which these models have often been incorrectly
compared:
The inter-relation between strain and stress is retained within the differential
equation, thus increasing numerical stability
The time consuming solution of the algebraic stress simultaneous equations is not
needed
This new model of eddy viscosity is therefore shown by Pope (1975) to correct the
following fundamental weaknesses of the Boussinesq stress-strain relationship:
Inability to capture normal stress anisotropy
Insufficient sensitivity to secondary strains
Excessive generation of turbulence at impingement zones
Violation of realisability at large ratios of strain
Unfortunately at the higher orders of expansion the model is restricted to two-
dimensional flows as the three dimensional form is so intractable as to be of no value
55
(Pope 1975). Fortunately Speziale (1989), Suga (1996) and Craft et al (1996) have
made revisions to the anisotropic eddy viscosity relationship to allow their application
to a wide range of three dimensional turbulent flows. Craft et al (1996) details both
quadratic and cubic expansions of the Boussinesq hypothesis as follows:

,
_

,
_

+ +

,
_

+
ij lk lk jk ik
k
t
C
ki
S
jk kj
S
ik
k
t
C
ij kl
S
kl
S
kj
S
ik
S
k
t
C
k
ij ij
S
t
u u
j i


3
1
3

2

3
1
1
+
3
2
' '
(5.10)
where:

,
_

i
j
j
i
ij
x
u
x
u
S (5.11)

,
_


i
j
j
i
ij
x
u
x
u
(5.12)
The cubic expansion takes the following form:
kl kl ij
S
k
t
C
kl
S
kl
S
ij
S
k
t
C
ij nl mn lm
S
mj lm il
S
lm il
k
t
C
kl
S
li kj
S
lj ki
S
k
t
C
ij lk lk jk ik
k
t
C
ki
S
jk kj
S
ik
k
t
C
ij kl
S
kl
S
kj
S
ik
S
k
t
C
k
ij ij
S
t
u u
j i
+
+

,
_

+ +

,
_

+ +

,
_

,
_

+ +

,
_

+
2
2
7

2
2
6


3
2
2
2
5

2
2
4

3
1
3

2

3
1
1
+
3
2
' '


(5.13)
The coefficients C
1
to C
7
have been derived from considering the prediction of the
stresses from a wide range of turbulent flows. The first term on the right in Equation
56
5.13 is the standard approximation. It can be seen that the addition of up to a further
seven terms makes the relationship a great deal more mathematically complex and
detailed and constitute the additional terms which are quadratic and cubic in the mean
velocity gradients (Wilcox 1994). Detailed tests by the authors mentioned have shown
that the quadratic model showed significantly improved predictions for the
reattachment length of the separated region behind the backward facing step, with
results similar in accuracy to that obtained using the differential stress model.
Furthermore, the non-linear quadratic model successfully predicted an eight vortex
secondary flow in the non circular duct, an occurrence usually only predicted by
second order closure models which are able to predict the individual Reynolds
stresses. The quadratic models are designed for problems in which flow anisotropies
are the distinguishing feature. The cubic model includes extra terms in the constitutive
relation between Reynolds stresses and strain rates so it is reported to give better
predictions in curved flows, for example, over curved surfaces including hills.
Therefore, it could be argued that the quadratic model should be sufficient to describe
the flow in a majority of wind engineering bluff body simulations.
The quadratic non-linear model proposed by Speziale (1987) is as follows:
( )
( )
ij mm ij E
ij mn mn mj im D
ij
i
j
j
i
t j i
D D l C
D D D D l C
k
x
u
x
u
u u

1
1
]
1

+
3
1

3
1
+
3
2
2
2
' '
(5.14)
where
1
1
]
1

+
i
j
j
i
ij
x
u
x
u
D

2
1
(5.15)
and the Oldroyd derivative term (
ij
D ) is as follows:
mi
m
j
mj
m
i
ij
m
m
ij
ij
D
x
u
D
x
u
D
x
u
t
D
D

+ (5.16)

2 3
2
k
C l (5.17)
57
This model, again based on the work of Pope (1975), differs slightly from the
quadratic model of Suga (1996) due to the appearance of Equation 5.16 which is an
approximation derived from the revised relationship to a visco-elastic medium, see
Speziale (1987) for further details. Lumley (1970) explains how there are striking
similarities between the mean turbulent flow of a Newtonian fluid and the laminar
flow of visco-elastic fluid. Visco-elastic fluids, often termed Rivlin-Ericksen fluids,
are used in the description of dilute polymer solutions. According to Speziale (1987)
these fluids can be similarly modelled by the non-linear generalisation of the
Boussinesq hypothesis.
Due to the nature of the non-linear model, comparisons have been made with the
second order closure, algebraic stress model. There is a similarity in both models in
that the stresses are linked non-linearly to all strain components, but beyond this the
two models are very different as the algebraic stress model is based on simplified
transport terms for the full Reynolds stress equations.
In a similar manner to the k and equations and the Reynolds stress transport
equations detailed earlier, the non-linear models are rigorously derived from either the
Navier-Stokes equations or the eddy viscosity Boussinesq hypothesis. The author is of
the opinion that turbulence models of this kind, rather than turbulence models
modified in an ad-hoc manner, are the most sensible way to improve the current
modelling technology.
Reported deficiencies and errors with these models are particularly scarce due to the
minimal testing undertaken in this field. Nevertheless it has been stated that the cubic
non-linear model is the preferred expansion due to the fact that it can be applied to a
greater range of flow fields, including curved surfaces (Craft et al 1996). As such the
model constants used to close the equations can be calibrated from a much wider
range of flows thus making the model more universally applicable than the quadratic
model. In addition, it should be noted that the algebraic expressions for the Reynolds
stresses are unable to model fluid transport effects, in contrast to the differential stress
model (see section 4.2.7).
Consequently as these models have not yet been applied to three dimensional wind
engineering flow fields the author has undertaken extensive testing and offered
critical discussion of the findings in chapter 6.
58
5.5 Wall Reflection Terms The Pressure Strain Relationship for the
Differential Stress Model
5.5.1 Theory
Referring to the work of Uzkan and Reynolds (1967), Thomas and Hancock (1977)
and Perot and Moin (1995) it is apparent that the effects of a solid wall on turbulence
behaviour are very complex. Viscous effects in the near wall region damp the
turbulence intensities, while wall blocking effects amplify tangential turbulence
intensities. This is due to the fact that the fluid cannot penetrate the boundary and as
such any fluid moving towards the wall must eventually turn and move parallel to the
wall. In this way energy is transferred from the normal velocity to the two tangential
velocity components increasing the turbulence intensity and anisotropy near the wall
(Perot and Moin 1995). Consequently the distinctive features in this region
immediately adjacent to a solid boundary are strong inhomogenity and excessive
turbulence anisotropy. This is a non-trivial effect due to the solid boundary that is
known as the pressure echo effect or wall reflection effect.
5.5.2 Modelling
In the differential stress model the pressure-strain term is important in explaining the
correlation between the fluctuating pressure and strain. In non-wall bounded flows
this term is used to derive the pressure-strain fluctuations due to interactions between
adjacent eddies and eddies and the mean flow and to steer the turbulence towards
isotropy. In wall bounded flows it is necessary to further predict the influence of the
wall on the flow field. Thus a pressure-strain term is required that is applicable to all
points in the flow from the wall region to an infinite distance from the wall. Adequate
mathematical expressions to describe the behaviour of the wall pressure reflections
and pressure strain interaction have proved particularly difficult to find. This is due to
the significant complexity of the problem and the existence of unmeasurable terms in
the pressure-strain term leading to highly empirical models. A great degree of
cleverness and ingenuity is required to establish a rational closure approximation. The
problems caused by the wall are further compounded by the use of isotropic
dissipation models, in the Reynolds stress transport equation, which imply dissipation
of energy by the small scale isotropic eddies. This is incorrect near the wall where the
59
eddies are small, i.e. no larger than the distance from the wall, but are highly
anisotropic due to the effects of the solid boundary.
Early Reynolds stress models, such as Gibson and Launder (1978), have used
isotropic or quasi-homogeneous assumptions in some important aspects of their
formulation. Such formulations tended to increase the isotropisation process as the
wall is approached thus leading to non-physical representations (Leschziner 1995). In
order to perform the near wall integrations these traditional pressure-strain terms
require the introduction of a variety of ad-hoc wall damping and wall refection terms
that depend inversely on the distance from the wall as well as the unit normal to the
wall. These damping functions essentially correct the erroneous behaviour of the basic
models by adjusting the model solution to fit a particular data set. The isotropisation
process, correctly predicted in non-wall bounded flows, is damped to allow the
turbulence to become increasingly anisotropic due to the influence of a solid
boundary. These case specific revisions are against the very nature of second moment
closure models (differential stress models) which are said to be more universally
applicable. Furthermore, the introduction of the damping terms makes the application
of these models to flows in complex geometries very difficult, as the local wall
distance is not uniquely defined. Consequently, at present, in order to use the DSM as
a generic wind engineering model, applicable to all geometries and simulations, it is
necessary to omit the wall reflection terms. The only known alternative, at present, to
this is to use non-linear pressure-strain terms that asymptotically follow the correct
distribution of stresses as the wall is approached. For example the SSG model
(Speziale, Sarker and Gatski 1991) does not use a wall reflection term in the pressure-
strain relationship, instead a number of the constants in the pressure strain term have
been calibrated with homogeneous shear flow. Unfortunately although these models
perform reasonably well in simple flow fields they show considerable errors when
applied to wind engineering flow fields. For example;
The use of non-linear models, together with an inhomogenity correction to

ij,2
, is not satisfactory for handling flows impinging on a wall (Gatski et al
1996, p273).
This non-linear scheme has so far been tested only for flows parallel to the
wall. Whether it will be satisfactory for flows impinging on a wall is open to
doubt (Launder and Li 1993, p1005).
60
Furthermore Speziale (1994), in a review of the most recent state of the art second
moment closure models with wall effects, stated that the move towards the non-linear
models is not without problems even in generally simpler flows without impinging. It
is apparent that at present there is no model that is generally applicable to a wide
range of flow fields and disciplines, consequently many researchers have produced
unworkable models that are overly complex and empirical.
An early example of a pressure-strain term with wall reflection effects, that has been
tested in chapter 6, is given below. The pressure-strain term used consists of the sum
of three parts as follows:
A return to isotropy term involving the fluctuating velocities only (Rotta 1972)
( )

ij i j ij
c
k
u u k
,1 1
(5.18)
A return to isotropy of the production term (the rate of production of the Reynolds
stresses) (Naot et al 1970)

ij ij ij
c P P
,2 2
2
3

_
,

(5.19)
The wall reflection terms to account for the effect of walls on the pressure-strain
terms (Gibson and Launder 1978)

st w n ij n i ij n j ni nn ij ni ni nj ni
n
c
k
u u u u u c f
L
x
, , , ,

_
,
+

_
,

_
,

_
,

1 2 2 2 2
3
2
3
2
3
2
3
2
(5.20)
where the index, n, indicates the direction normal to the wall. The function f takes
the form:
f
L
x
C
k
x
n
w
n

_
,

15 .
(5.21)
61
where x
n
is the normal distance from the wall. As these terms are a simplification of
more complex relationships experimental constants are incorporated to simplify the
equations.
The interested reader is referred to Bernard and Speziale (1992) for details of future
proposals for improved pressure-strain terms.
In a similar manner to the two layer models described earlier, full and detailed wind
engineering testing of a number of the wall reflection models, described above, has
already been undertaken for a surface mounted cube in an atmospheric boundary
layer, the interested reader is referred to Murakami et al (1993). Further details are
given in chapter 6.
5.6 Direct Numerical Simulation (DNS)
This technique is the easiest to define but is the most computationally expensive.
Direct numerical simulation solves the instantaneous Navier-Stokes equations for the
full range of turbulent motions from the largest scales down to the dissipative scales.
As such DNS does not involve turbulence modelling of any kind, the flow field is in
fact discretised corresponding to the Kolmogorov microscale. This approach yields
the complete spatial and temporal state of the turbulent flow. Full details of
experiments conducted using DNS can be found in Le et al (1993).
The overwhelming problems with this technique are the massively high computing
costs and memory requirements involved. An estimate of numerical grid and CPU
requirements for predicting high Reynolds number flows is shown in Table 2.
Reynolds
number
6600 20,000 100,000 10
6
Grid nodes 2 x 10
6
40 x 10
6
3 x 10
8
15 x 10
12
Time at 150
Mflops
37 hours 740hours 6.5years 3000years
Time at 1
Tflops
20secs 400secs 8.3hours 4000hours
Table 2: CPU requirements for DNS (Leschziner 1995)
62
An excellent example of the massive computational requirements is given by Speziale
(1991) who states that a typical flow domain of 0.1m by 0.1m with a high Reynolds
number turbulent flow would contain eddies down to 10 to 100m in size. These very
small eddies would naturally have very high frequencies of approximately 10kHz,
thus requiring time steps of about 100s. In order to directly capture the details of the
smallest eddies it has been calculated that a computational finite volume mesh of 10
9
to 10
12
points would be needed. The direct numerical simulation of such a turbulent
flow at a Reynolds number of 500000 would require a computer 10 million times
faster than a current generation supercomputer.
Due to the extremely high computational costs incurred, DNS is not able to calculate
anything other than low Reynolds number flows in simple geometries. The simplest
application of DNS requires access to a supercomputer. It is apparent that more
practical applications of this technique await significant developments in computer
technology. Nevertheless at present DNS has proved invaluable in supplying
computed statistics that can be used to test proposed closure approximations in
engineering models. DNS has been used at a fundamental level to obtain a greater
understanding of turbulence structure and processes that are of value in developing
turbulence theory that would be impossible to measure with traditional experimental
techniques, for example the dissipation rate of turbulence (Lee et al 1990).
5.7 Large Eddy Simulation (LES)
This technique directly resolves the large eddies present in turbulent flows and models
the smaller scale eddies. LES rests on the supposition that the most important
turbulent transport processes arise from the large to medium scale eddies while
smaller eddies are principally responsible for dissipation of the turbulent energy.
Direct numerical simulation is used to resolve the eddies that are larger than the size
of the finite volume cell, while a RANS or simpler, model is used to model the more
universal nature of the small scale eddies that are smaller than the mesh size. Since
Reynolds averaging is not applied to the basic equation used in LES the scale of the
resolved eddies is much smaller than when using, for example, the k- model. The
main task of this so called sub-grid scale model is to extract the appropriate amount of
energy from the resolved flow field and dissipate it (Murakami et al 1987). As LES
models the small scale turbulence a much coarser mesh, and larger time-steps, can be
used than in direct numerical simulation in which eddies are resolved to the
Kolmogorov microscale. Furthermore, as the size of the smallest eddies decreases
63
with increasing Reynolds number it is possible, for a given computing cost, to achieve
simulation of much higher Reynolds number flows with LES than DNS.
The size of the grid cell used in large eddy simulation generally defines the level at
which the smaller scale eddies are filtered and thus modelled. For this reason this
technique is often referred to as space filtering.
5.7.1 Discussion.
This technique, vigorously pursued in Japan and the USA, has the advantage of
producing time dependent flow information of generally very high quality and
accuracy even in complex flow fields such as those found in wind engineering. It is
widely felt that LES is a promising tool for the future:
LES has succeeded in reproducing the properties of a highly anisotropic flow
field in wind engineering problems (Murakami and Mochida, 1989a, p69)
Present difficulties in using LES mainly revolve around the constraints on available
computer processing time and storage capacity which effectively hold back its use and
advancement. This technique, although being more economical than DNS is still very
resource intensive and as such is not yet used outside of the research community.
Further difficulties are apparent in the use of the sub-grid scale model to dissipate the
flow energy. Leschziner (1993) reports that the sub-grid scale eddies are not simply
dissipative but can contribute significantly to turbulent mixing. Such processes as
backscatter can occur in which the small eddies combine with larger eddies and
transfer energy to them.
A number of models are available to be used as sub-grid scale models ranging from
simple Smagorinsky eddy viscosity models to full Reynolds stress closure models.
Due to the promising results from this method it was considered important to further
evaluate the necessary computational requirements of this technique, consequently
further details and experimental testing are discussed in chapter 7.
64
5.7.2 Coherent Structure Capturing (CSC), or Very Large Eddy Simulation
Coherent structures are usually defined as the large turbulent structures observed in
many shear flows. The accepted features of these structures are their large size and
low frequencies such as the horseshoe vortex (see section 2.2.2) and the large vortices
produced in the wake of a bluff body (Ferziger 1994). It has been found from
experiments that these coherent structures account for a significant fraction of the
turbulent kinetic energy and Reynolds stresses and fluxes in turbulent shear flows.
Many flow fields, particularly those found in wind engineering, contain these large so
called coherent structures that can in many ways define the flow field. If LES is
applied to flow fields with a small number of these energetic coherent structures
which are larger than the bulk of the motion that constitutes the turbulence it is
possible to increase the size of the filter. This effectively increases the coarseness of
the grid, retaining the large-scale coherent structure while omitting more of the small-
scale turbulence. The theory being that the coherent structure is responsible for more
of the turbulence behaviour thus further reducing the importance of the smaller scale
eddies.
Ferziger (1993) gives an example of the reduced mesh size required to calculate the
flow over a circular cylinder.
Technique D.N.S L.E.S C.S.C
Grid points 10
12
10
10
- 10
8
10
6
Table 3: CPU requirements for CSC (Ferziger 1993)
5.7.3 Discussion.
As the success of this method relies on the simulation of large, essentially coherent
structures responsible for the large scale pressure field rather than eddies responsible
for momentum and energy transport it is felt that CSC may be well suited to wind
engineering applications (Ferziger 1993). As such this technique has moved the use of
LES one step closer to mainstream application in the field of wind engineering.
Unfortunately, although CSC has low computational costs in comparison to DNS and
standard LES, it is still a relatively computationally intensive procedure. Furthermore,
the use of CSC in LES requires considerable expertise in constructing suitable grids
65
that can capture the correct detail. Further detail on this subject is given by Speziale
(1998).
5.8 Summary
The literature search conducted thus far has very quickly highlighted the deficiencies
in the most popular turbulence models, including the errors in the isotropic eddy
viscosity concept and the possible numerical instabilities found when applying the
DSM to complex wind engineering flow fields. Although full testing was required
before definitive conclusions could be drawn, this information was very useful in
giving direction to the project and demonstrating the requirements for improved
turbulence models for computational wind engineering. It is apparent that a turbulence
model, suitable for wind engineering applications, should be able to model anisotropic
turbulence, as the DSM, but it should also be sufficiently numerically stable to allow
its application to the vast majority of flow fields, as the standard k- model. These
early investigations suggested that a turbulence model is required that incorporates the
best aspects of both these turbulence models, such as the non-linear k- models
detailed in section 5.4. The fact that these models may offer potential improvements
and as yet have not been fully applied to the field of wind engineering made them an
ideal subject for this thesis. It is also clear that the technique of very large eddy
simulation is becoming of greater importance as computers become ever more
powerful. Therefore the following chapters will address and discuss these points.
66
6 Results: A Comparison of Data Obtained from Both CFD
Simulations and Full Scale Experimental Studies.
6.1 Introduction
Originally this project proceeded by testing the accuracy of the more popular
turbulence models detailed in chapter 4, thus determining the validity of the
information found in the literature search when the results were compared to full-scale
rather than wind tunnel scale buildings. The outcome of these results suggested that
the early considerations regarding the standard k- and differential stress models were
in fact correct. Further tests were then completed using a number of the models
detailed in chapter 5.
The following chapter will therefore detail all of these tests together to enable direct
comparison of the results obtained from a total of eight turbulence models. As far as
the author is aware these tests are the first to detail comparative results of CFD and
experimentally obtained data for a full-scale surface mounted cube. In addition, it is
the first time the non-linear and RNG k- models have had their predictive accuracy
assessed when applied to wind engineering flow fields.
Overall results and conclusions will be given in both chapters 6 and 8 for all models
tested in this project based on the following criteria, which the author considers to be
the essential requirements for a suitable wind engineering turbulence model:
overall accuracy
stability and ease of use in complex flow fields.
computational requirements and overheads.
Although the accuracy of the model is the prime consideration, the model
applicability is also of paramount importance. An accurate turbulence model that
cannot be applied to say 6 out of 10 flow simulations, due to poor numerical stability,
is of little use to the computational wind engineer. Steady state simulations using
modern computer facilities makes the issue of computational overheads only a
relatively minor concern, nevertheless this point will be addressed as fast simulation
times are preferred. Finally it is also preferable that the model has a general
universality and so can be applied with equal success to other engineering disciplines
and thus gains acceptance among the engineering community in general.
67
In order to assess the relative abilities of the turbulence models these tests included
comparisons of 1:1 scale CFD and experimentally obtained data for a 6m surface
mounted cube both normal and skewed at 45 to the incident wind. Further tests have
included the Silsoe structures building at normal and transverse orientations and the
Silsoe experimental wall, all of which are detailed later. Validation tests were carried
out using the two dimensional backward facing step. Test variations included the use
of different turbulence models, inlet velocities and convective differencing schemes in
order to ascertain their effect on the accuracy of the flow field.
6.2 Full Scale Test Details
6.2.1 The 6m Cube
A 6m cube, located at the Silsoe Research Institute (SRI), was recently constructed to
allow the comparison of full-scale rather than model scale wind tunnel data to CFD
derived results. The cube is positioned so that the boundary layer is generated from a
fetch consisting of short grass with an effective roughness length of 0.01m. Checks
undertaken in previous years have shown that the effective roughness length of the
fetch is constant due to regular cutting of the grass (Quinn 1999). The cube can be
rotated through 360 and pitched on the horizontal axis by 5. The cube surface
consists of sheet metal cladding with a smooth plastic coating to afford protection and
avoid changes of the surface roughness due to rust. A photograph of the cube can be
seen in Figure 14.
A number of measurements were made of the pressures around the cube and the wind
dynamic pressure at the roof height. The 16 pressure taps used were located along the
centre line of the cube on the windward face, roof and leeward face. The information
required to derive the pressure coefficients was obtained from an ultrasonic
anemometer positioned 25m upstream of the building at roof height. The pressure tap
locations were also used to obtain data when the cube was skewed at 45 to the
incident wind. The tapping points were constructed of simple seven millimetre
diameter holes (a size sufficient to prevent water blocking the tapping points) and the
pressure signals were transmitted pneumatically, using a 6mm internal diameter
plastic tube to transducers mounted centrally within the cube. Tube lengths of up to
10 metres were used in this system giving a frequency response of 3dB down to 8Hz
(Hoxey et al 1999). The full recorded information consisted of simultaneous
68
measurements of all the pressures and the three components of wind at a sampling
rate of 5hz, which were processed to give all the required experimental data including
mean, fluctuating and spectral properties. A polynomial fit has been used to enable
actual measurements of wind speed and direction to be manipulated to give a full
quasi-steady prediction of surface pressure. These estimates are compared with actual
measurements in the form of statistical moments, probability densities and power
spectral densities. Mean pressure coefficients were derived from a recording period of
ten minutes (Hoxey et al 1999).
6.2.2 The Silsoe Structures Building
The Silsoe structures building (SSB) (Robertson and Glass 1988) was constructed
during 1986/87 to enable the measurement of full-scale wind loading and to aid the
understanding of gaseous and particulate pollution dispersion from agricultural
buildings. In order to further study the air flow characteristics around the building the
SSB was constructed with optional eaves geometry of the traditional sharp eaves or
curved eaves of 635mm radius. The 24m long by 12.9m span by 5.3m ridge height
building with a 10 duo-pitch roof (see Figure 15) is located on a flat, exposed site
adjacent to the 6m cube described earlier in this work (Hoxey et al 1995).
For this particular study there were a total of 32 pressure taps installed along the
various measurement lines, sampling at a rate of 5Hz, to allow a detailed study of the
pressures around the building with various incident wind angles. In a similar manner
to the 6m cube tests, a three component ultrasonic anemometer was used to allow
determination of the dynamic pressure upstream of the building at eaves height.
Records of one hour duration were made, and these were partitioned into 6 minute
records for analysis. All the pressures were measured relative to static pressure sensed
by a probe positioned upstream of the building where the influence of the pressure
field associated with the building was small. The pressure transducers were
automatically zeroed using a solenoid valve to equalise the pressure across the
transducer. This was followed by a calibration, again activated by solenoid valves,
where the wind dynamic pressure was applied to all transducers (Hoxey et al 1995).
Richardson et al (1995) has stated that this procedure ensured a pressure measurement
accuracy of better than +/- 1N/m
2
.
69
All the experimental data was processed to give mean, rms, maximum and minimum
values of surface pressures and similar statistics for the wind dynamic pressure,
together with the mean and rms of the wind direction. Further details of the tests are
given by Richardson et al (1995) and Hoxey et al (1995) and Robertson and Glass
(1988).
6.2.3 The Silsoe Experimental Wall.
The Silsoe experimental wall and the various studies associated with it are described
in detail by Robertson et al (1996, 1997). The information below is an extract from
one of those papers used to inform the reader of the main technical details of the wall
and the relevant testing procedures.
The experimental wall (shown in Figure 16) has a height of 2m and is 215mm thick
(representing typical freestanding masonry walls). It was erected on a specially
prepared concrete foundation but, for experimental convenience, was constructed not
of bricks but of 2m high steel posts at 2m centres and square, hollow, timber panels
which were located between pairs of posts. The panels were readily removable and
relocatable, enabling the length of the wall to be changed from 2m (one panel) to 18m
(nine panels) in 2m increments. At either end of the 18m long wall, a corner could be
formed as required by introducing one or two wall panels to form a perpendicular run
of wall of 2 or 4m in length. The wall was constructed on the flat and exposed site at
Silsoe, near the Silsoe structures building and the 6m cube. One of the 13 panels
contained 15 tappings on each face located at corresponding positions in three rows of
five columns. All the pressure transducers, switching solenoids and power supplies
were mounted inside the panel. An ultrasonic anemometer, static probe and
directional pitot tube were mounted at a height of 2 metres on a reference mast
positioned some 15 metres off one end of the wall. When the wind speed and
direction complied with the prescribed settings, the data logger automatically
recorded at 5Hz the outputs from the 30 pressure transducers, the load cell and the
ultrasonic anemometer for one hour. The pressure transducers were automatically
zeroed and calibrated at the start and end of each hour using the pitot and static probe
pressures (Robertson et al 1996, 1997).
70
6.3 CFD Test Details - Accuracy Checks
6.3.1 Introduction
In any numerical simulation of fluid flow there are a number of checks that have to be
made to assess the overall accuracy of the results obtained and to ensure that the CFD
data is of a sufficiently high quality. The following work details all these important
checks and wherever necessary explains any assumptions that have been made.
The CFX-Visualise package was used to obtain velocity, pressure and turbulent
kinetic energy profiles for the cube, SSB and Silsoe experimental wall to be used in
the following checks and comparisons. Linear interpolation between the nodal points
with the CFX post processor meant that a minimum of 11 data points were recorded
for each metre length of the domains for use in the various graphs in the following
work. In order to obtain the pressure data the measurement lines were located
approximately 10mm from the face of the various building shapes. To obtain velocity
and turbulent kinetic energy profile data the measurement lines were located
approximately 50mm from the respective faces. Further detail on the exact positions
of the measurement lines used for each test case is given in sections 6.5.1, 6.5.5, 6.6
and 6.7. All vector and contour plots shown are located along either the vertical or
horizontal centrelines of the cube, SSB and experimental wall.
Unless otherwise specified this section refers equally to all the models tested.
6.3.2 Use of Full-Scale Results for CFD Validation.
Although not strictly a numerical accuracy check, there are a number of accuracy
issues that are very important to this project when using full-scale rather than wind
tunnel scale building models. This sub-section will briefly highlight some of the more
important aspects with regards to this project, for detailed information the interested
reader is referred to Cook (1986a, 1986b).
For the full-scale observations, the most apparent advantage is that they do not suffer
from any scale mismatch due to Reynolds number, wind shear and turbulence
intensities or from blockage effects. On the other hand, they are costly and time
consuming. It is, moreover, impossible to control the approach flow conditions, e.g.
71
the static and dynamic reference pressures, the thermal stratification etc., which will
inevitably obscure details in the observed data (Holscher and Niemann 1998). In
addition, it is stated that the error of measurements could possibly be larger for full-
scale observations compared to well controlled wind tunnel tests, which should be
borne in mind when using them for calibration, such as is the case for this project.
Nonetheless there are serious matters to be considered when undertaking scale
simulations in a wind tunnel. One of the most important aspects, with regards to this
project, are the effects of the differing Reynolds numbers involved in the areas of
flow recirculation over the cube models. For the skewed cube case experimental
evidence shows the existence of delta wing vortices over the roof. These are regions
of flow involving very high re-circulatory velocities (Hoxey 1998). The scaling down
of such regions of flow to wind tunnel models will result in very small vortices that
will be influenced to a much larger degree by the effects of the fluid viscosity, thus in
effect reducing the strength of the predicted vortex. These so called Reynolds number
vortex effects may result in unrealistically small roof and wake recirculation zones,
for both the normal and skewed cube cases, and thus errors in the prediction of
pressures. Therefore the novel use of 1:1 scale CFD and experimental models should,
in theory, eliminate this detrimental effect.
Referring to Davenport 1999:
In this paper we have stressed the importance of full-scale testing as a
benchmark for both wind tunnel and CFD modelling. As Martin Jensen has
reminded us, the lack of full-scale verification that has been tolerated is
embarrassing. It is not characteristic of other technologies such as shipping,
transportation or aeronautics. (Davenport 1999, p2)
Finally in this section, it is important to discuss the probable error in the mean
pressure coefficient data, used in this thesis, for the full-scale cube, SSB and
experimental wall tests. Hoxey (1999) and Quinn (1999) have stated that the predicted
error for the individual RMS readings for a full-scale test with a ten minute sampling
time were approximately t 4 percent. Furthermore, by repeating the tests the
individual reading error is reduced (ten minute sampling time error divided by the
square root of the number of times the test is repeated). As these tests were repeated
for ten samples the approximate error for the individual mean pressure readings was
reduced to between t 1 and 2 percent, although minor changes in the incident wind,
between samples, will have an affect. This predicted error was used, as a general rule
72
of thumb, for all the full-scale data used in this project. Nonetheless, it is clear that
this error will also be dependent on the location of the pressure tap, with the above
mentioned error in an area of steady flow. The error may marginally increase in areas
of unsteady flow, such as in the wake (Hoxey 1999). Further discussion on the
experimental errors in the full-scale tests is given later in this chapter.
6.3.3 CFD Solver
As previously explained all tests for this project made use of a commercially available
code, CFX (CFX-International 1998). This route was taken as opposed to the
development of a bespoke code for several reasons:
CFX, like other commercial codes, has been validated for many test cases in a
number of different application areas.
The tools for geometry building, mesh definition and the presentation and analysis
of the results are versatile and readily available.
The code has many options for different methods and solution algorithms.
Transfer of results and recommendations to end-users is easily accomplished as
the code is available for many computer platforms and there is comprehensive
documentation and training.
Rapid development times for new model implementation.
Where necessary use was made of the ability to add user-defined FORTRAN
subroutines to accomplish specific tasks and for more substantial alterations,
including the addition of new turbulence models, the University of Nottingham had
access to the source code for CFX4.2 (Wright and Easom 1999).
6.3.4 Geometry
Although the tests undertaken in these simulations compare the CFD and full-scale
experimental data at an equal scale, it is nevertheless particularly important to specify
the correct boundary conditions and domain dimensions for the CFD runs, as is the
73
case in actual wind tunnel experimental studies. Therefore the sides, lid and most
importantly the inlet and outlet of the domain were positioned so as to minimise the
interference with the flow field around the cube. The height and width of the domain
was decided upon after discussions with Professor Baker (1997) and reference to
Baetke et al (1990). Baetke et al (1990) states that the blockage ratio, defined as the
ratio of the frontal area of the cube to the vertical cross sectional area of the
computational domain, should be no greater than 3 percent. The blockage ratio for
these tests ranged from approximately 1.5 to 3 percent. Furthermore, symmetry
boundaries were specified to further reduce the effect of the sides and lid of the
domain. The conditions at this boundary are:
1. no flow across the boundary
2. no scalar flux across the boundary.
The normal velocities are set to zero at the symmetry boundary and the values of all
other properties just outside the solution domain are equated to the nearest node just
inside the domain.
The position of the outlet boundary is another important consideration (see Figure
17). The outlet boundary for a flow simulation must be placed at a sufficient distance
from the inlet and obstacle to the flow so that there are no flow gradients in the flow
direction. The flow at outlet must be in a state of equilibrium otherwise the interior
solution to the flow problem will be influenced and the accuracy of the solution will
be compromised. For high accuracy solutions it is recommended that the flow
solution be compared for a number of different outlet positions to test sensitivity
(Versteeg and Malalasekera 1996).
The domain was meshed to allow a high degree of resolution of the flow with
particular attention paid to areas of high flow gradients at the points of flow
impingement and separation. Full details on the grid refinement tests are given in
section 6.3.9.
The computational domain for the skewed cube, whose relative length and width are
equal to 2 (6 cosine (45)), was constructed to adhere to the blockage ratio
requirements previously discussed.
74
6.3.5 Atmospheric Boundary Layer Inlet Conditions.
As the main aspect of this project was to determine the accuracy of a number of
turbulence models in wind engineering flows it was particularly important to remove
any other differences between the experimental and numerical results. It is well
known that in a numerical wind engineering simulation even minor changes to the
inlet conditions can significantly modify the predicted flow field. It is of paramount
importance that the conditions at the inlet of any CFD simulation match, as far as
possible, those of the wind tunnel or full-scale study. In order to obtain CFD results
that could be directly compared with mean experimental values the ground roughness
length was set equal to 0.01m to represent the fetch at the SRI and the cube walls
were given a roughness length equal to 0.005m. Fully developed equilibrium
atmospheric boundary layer flow profiles including variables such as stream-wise
velocity, turbulent kinetic energy and Reynolds stress were specified at the inlet.
These profiles allowed for a sustainable equilibrium boundary layer when used in
conjunction with the appropriate wall roughness lengths. Full use was made of user
FORTRAN routines that allow the user to override the simple default settings in CFX.
6.3.5.1 Silsoe Research Institute: Boundary Layer Inlet Conditions for the k-
Turbulence Model
Richards and Hoxey (1993) and Richards (1989), detail the computational inlet
conditions for the Silsoe fetch for the k- turbulence model. The following work
details the methods used in deriving the appropriate equations.
In modelling the atmospheric boundary layer there are a number of important rules to
follow that have been defined by Jensen (1958) and Ludwig and Sundaram (1969).
The classical approach, and the starting point, is to define a two-dimensional
thermally neutral boundary layer in which Coriolis forces are ignored, with all flow
variables in equilibrium. It is stated that in steady incompressible two dimensional
flow modelling the existence of homogeneous flow has the following properties:
1. the vertical velocity is zero
2. the pressure is constant
3. the shear stress is constant.

i.e.
2
*
) ( u
z
u
O t l

+ (6.1)
75
Using these rules and assuming that the atmospheric boundary layer is in equilibrium
with the ground, the researchers at the SRI used a suitable velocity profile at the inlet
based on the Harris and Deaves (1981) model. This model states that the mean
velocity profile in the lower part of the atmosphere may be adequately described by:
1
1
]
1

,
_



z
z
z z u
z U 75 . 5 ln ) (
0
0
(6.2)

,
_

,
_

0
ln
.
z
Z
U
u
ref
ref

(6.3)
(Richards 1989)
The corresponding derivation of the k and profiles is obtained from Richards (1989).
In the equilibrium boundary layer it is assumed that the shear stress decreases
parabolically with height, as follows:
2
2
) (
1
1
]
1

,
_


z
u
z
(6.4)
As an equilibrium boundary layer is being modelled and there is no change of fetch
the rate of generation of turbulent kinetic energy will ideally equal the rate of
dissipation. This further implies that:
( )
( ) ( )
z
u
k
z z
z

. = of generation (6.5)
So substituting Equations 6.2 and 6.4 into Equation 6.5 gives
( )
( )
( )

,
_

,
_

+
+
+
1
]
1

0
0
2
2
75 . 5 1 . 1
z z
z z
u z
u
z
(6.6)
Now turning the attention to the complementary profile equation for the turbulent
kinetic energy k:
76
( )
2 2 2
2
1
w v u k + + (6.7)
The velocity fluctuations or normal Reynolds stresses in this equation may be derived
by considering the turbulent viscosity as defined in the k- model:



2
k
C
t
(6.8)
Referring back to chapter 4, because the turbulent viscosity greatly exceeds the
laminar viscosity it follows that:
( )
( )
z
u
z
t z

(6.9)
and hence from Equations 6.3, 6.4, 6.7 and 6.8:

C
z
u
k
2
2
1
,
_

(6.10)
In many cases the domain is considerably shallower than the layer thickness, or so
called gradient height, which is typically at least a few hundred metres. In such
circumstances it is reasonable to set the function of variation of shear stress across the
layer equal to 1 (Richards and Hoxey 1993).
Equations 6.2, 6.6 and 6.10 are reduced accordingly as follows:
( )
ln
0
0

,
_


z
z z
k
u
U
z
(6.11)

C
u
k
2

(6.12)
( )
0
3
z z
u
+

(6.13)
77
Equation 6.11 has been derived from the Deaves and Harris (1981) mean velocity
profile model, as stated earlier. Further assumptions regarding the shear stress
profiles, the eddy viscosity concept and equilibrium relationships between kinetic
energy production and dissipation have resulted in Equations 6.12 and 6.13.
Theoretically these equations should represent the boundary layer profiles of the
atmospheric wind at the SRI. Richards and Hoxey (1993) have directly compared
these profile equations with full-scale site data and found that there is an acceptable
match. Modifications since 1993 have included the inclusion of acceptable decay
profiles with height for both the k and equations. Furthermore, these equations have
been accepted by CFX International as suitable representations of the boundary layer
and are supplied to users upon request (Sinai 1995).
6.3.5.2 Appropriate Inlet Conditions for the Differential Stress Turbulence
Model
A further procedure was necessary to generate equilibrium profiles of variables for the
differential stress model that could be entered into the cube flow field simulations as
boundary layer inlet conditions. These tests used a short length of a computational
domain with mass flow boundaries at inlet and outlet set as a single patch with zero
mass flux. This method effectively recycles the outlet flow therefore replicating a
very long wind tunnel or length of fetch. As it is not possible to set an inlet velocity
using this method, the floor of the virtual wind tunnel was made to move with a
horizontal velocity. The moving floor ensured that the necessary shear in the flow was
generated due to the specification of a ground roughness equal to 0.01m. Using this
technique it was envisaged that the turbulent kinetic energy, Reynolds stresses and
energy dissipation profiles would develop in the same way as if an inlet velocity had
been specified, as these flow variables are invariant under coordinate transformation.
These tests were also undertaken for the standard k- model.
The boundary layer data generated was incorporated into the CFD simulations by
fitting curves to the first 40m of the profiles and entering the equations into the
appropriate user FORTRAN routine.
Validation of the numerical boundary layer generated was possible through
comparisons to published work by Richards and Hoxey (1993) entitled Appropriate
boundary condition for computational wind engineering models using the k-
78
turbulence model. This paper details equations that specify the velocity profile as
well as a maximum value of turbulent kinetic energy and energy dissipation at ground
level for the Silsoe fetch (see section 6.3.5.1). Figure 18, which shows velocity
profiles for the differential stress and standard k- models, with a reference velocity of
10m/s at 60m, shows that there is an exact match between the numerically derived
velocity profiles from the SRI data and the DSM numerically derived profile.
Substitution of the friction velocity into the turbulent kinetic energy equation derived
by Richards and Hoxey (1993) shows that the differential stress model accurately
predicts the maximum ground turbulent kinetic, at 0.00008m
2
/s
2
for a simulation
velocity of 0.1 m/s at 60m (as shown on Figure 19). It is explained by Sinai (1995),
that in the absence of information on the shear stress profiles a constant, linear or
more accurately a quadratic decay profile should be assumed. Analysis of the profile
shown in Figure 19 shows that the quadratic profile assumptions are in fact closest to
the decay profile generated by the computational simulations. Therefore using the
available experimental boundary layer data it is apparent that the DSM predicts an
accurate match to experimentally obtained mean data. In addition, previous
investigations have been undertaken to obtain more detailed information on the
boundary layer at the Silsoe Research Institute. To date a number of experiments have
been conducted to provide full-scale fluctuating velocity data using both 20hz and
100hz anemometers (Hoxey and Richards 1992 and Richards and Hoxey 1999). Mean
profile data from these results for the various Reynolds stresses have highlighted
discrepancies that, for example, show the magnitude of the stresses increasing with
height above ground. It appears from the spectral analysis that the anemometry
equipment was unable to record the correct magnitude of the Reynolds stresses due to
the smaller scale higher frequency fluctuations near the ground, particularly in low
wind speeds (Quinn 1999). This is perhaps an unfortunate consequence of using full-
scale comparative data. It has therefore proved difficult to use any of the more
detailed experimental data thus far. Nonetheless it should also be emphasised that
simulations with the DSM have used sustainable equilibrium boundary layer inlet
conditions generated as explained. Any attempt to force any other inlet conditions
would result in a rapid decay or change in the variable profiles as the computational
fetch is traversed.
6.3.6 Variations on the Ground and Building Roughness Length
Discussions with Drs Hoxey (1999) and Robertson (1999) at the SRI suggest that the
roughness length of the fetch at the Institute may vary, throughout the year, from
79
approximately 0.008 to 0.04m. It was also stated that variations between these limits
have no noticeable effect on the recorded full-scale mean pressure distributions. The
majority of CFD tests undertaken have used an average roughness height of 0.01m. It
was therefore decided to undertake a small number of tests with modifications to the
roughness height between the limits discussed. Figures 20 and 21 show the windward
face and roof pressure distributions for the quadratic non-linear k- model, firstly with
the ground and inlet roughness length set to 0.04m and secondly with the ground
roughness at 0.01m and the building roughness reduced to 0.001m. Interestingly it can
be seen that for the computational tests the increase in the turbulence due to the higher
roughness length has the effect of marginally increasing the front face stagnation
pressure and increasing the roof front corner maximum negative pressure. This in turn
causes a shorter recirculation zone that reattaches half way along the roof and a
steeper pressure gradient recovery (these effects are discussed in full in the following
chapters). These effects, whereby the flow and pressure fields are modified as the
incident wind turbulence increases, are well documented, the interested reader is
referred to Castro and Robbins (1977). The effects of changing the building roughness
were negligible. Drs Hoxey (1999) and Robertson (1999) of the Silsoe Research
Institute have confirmed that the cube, wall and SSB experiments were conducted
during the winter months. Furthermore analysis of the incident wind, by members of
the SRI, during these tests has further confirmed that the effective roughness length
was equal to 0.01m.
6.3.7 Revision to Wall Functions for the High Reynolds Number Turbulence
Models.
Analysis of the high Reynolds number wall functions used to model rough walls, and
included in the CFX supplied user FORTRAN, showed a discrepancy upon
comparison to the terms used in a number of other codes, including Star CD and
Fluent (Hargreaves 1998).
The standard wall function used in CFX is as follows:
) ln(
1
+

Ey
u
u
p

(6.14)
where u
p
is the velocity of the fluid at a point p near the wall, E is an empirical
constant, is the Von Karman constant, u
*

is the friction velocity and is given by:
80

w
u

(6.15)
where
w
is the wall shear stress.
When a rough wall is encountered with an average roughness height K
s
, Equation 6.14
needs to be modified in the following manner:
) ( ) ln(
1
s
p
K B Ey
u
u
+ +

(6.16)
where B is the amount by which the rough wall velocity profile changes and is a
function of the dimensionless roughness, . /

+
u K K
s
s With some manipulation
(section 14.6 of the FLUENT Users Manual Version 4.22 (1992)) this can be written
as:
)
3 . 0 1
ln(
1
s
p
K
Ey
u
u
+
+

(6.17)
with
) 3 . 0 1 /( ' s K E E
+
+ (6.18)
It is explained by Versteeg and Malalasekera (1996) that the thickness of the laminar
sub-layer will change depending on the effective Reynolds number at the near wall
node. In a similar manner the log layer thickness changes depending on the flow
velocity, Reynolds number and wall roughness along any given wall.
The standard CFX rough wall function incorrectly assumes a constant log layer value
as detailed in Equation 6.14 and therefore does not account for the changes described
by Versteeg and Malalasekera (1996). As a consequence it was necessary to amend
the rough wall functions used in CFX to allow calculation of a dynamic log layer
constant through the appropriate user FORTRAN routine and check a number of the
simulations to assess if there are any differences in the results obtained. Figures 22
and 23 show comparisons of the computed windward face and roof pressure
distributions predicted by the non-linear k- model, for the 6m cube, with CCCT
convective differencing. It is apparent from these results that there is a very small
81
difference between the use of the original and revised wall functions. This test was
repeated for a number of other turbulence models and cube orientations, which all
showed similar results. Nonetheless, all tests have used the revised wall functions
described.
6.3.8 Assessment of the y
+
Values
As explained in section 4.3 a near wall flow is taken to be laminar if the non-
dimensionalised wall distance
+
y is less than 11.22. When using high Reynolds
number turbulence models it is very important to check that the finite volume mesh
does not encroach into this region. If this is the case the turbulence model will attempt
to solve a transport equation validated for the turbulent region of the flow in an area
of purely laminar flow and consequently significant errors occur. As the simulations
presented in this thesis required high levels of accuracy it was necessary to use a large
number of nodes to grid the domain, therefore checks had to be undertaken to assess
whether the
+
y values were satisfactory. CFX4.2 can provide the
+
y values for every
node in the finite volume domain. As a consequence of these checks model domains
and finite volume meshes were amended as necessary.
6.3.9 Grid Refinement tests
The level of grid refinement can have a significant effect on the accuracy of the flow
pattern produced by the CFD simulation. In regions of separated recirculating flow,
such as found on the windward edge of the roof on a surface mounted cube, it is
possible to not resolve the change of direction of flow if a coarse grid is used. The
absence of separation is due to the integration of the flow over a whole cell, which if
only a small part of it is negative or reversed, will result in a net flow which remains
in a stream wise direction.
All CFD simulations should therefore be able to adequately prove that a sufficiently
fine mesh has been used. This is done by demonstrating that successive grid
refinements do not change the flow field in any way. Therefore extensive grid
independence checks were undertaken with both the cube and SSB for a number of
model orientations as detailed below. Simpler tests were carried out for the
experimental wall structure, which was found to be less sensitive to grid refinements.
The grid independence tests were as follows:
82
1. a 300 percent increase in the number of grid points in the computational domain
(total nodes approximately 550,000).
2. As part 1, plus a further 200 percent to 450 percent increase in the number of grid
points around the cube only (i.e. increases from 20
3
to 90
3
grid points on the
actual cube), depending on the cube orientation (total nodes 850,000 to
1.2million).
3. For the SSB three successive grid refinement tests were undertaken with,
approximately, a 100 percent increase, per refinement, in finite volume nodes,
using the most critical test case, the SSB with sharp eaves.
Test part 2 was necessary in order to confirm or otherwise that small scale flow
effects, such as the delta wing vortex on the skewed model, had been adequately
resolved. Obviously this is particularly important in order to accurately predict the
pressure distributions and the cube flow field. Murakami (1990) further discusses the
above mentioned points. All these tests used the RNG and non-linear k- models and
CCCT differencing.
Figures 24 and 25 show the RNG k- model predicted windward face and roof
pressure distributions for the normal cube test 1. It can be seen from this graph that
there is a minimal change between the two grids used. These results were also found
for other turbulence models including the non-linear and standard k- models and for
the skewed cube orientation.
Using only the normal cube, test number 2 showed revised pressure distributions for
the roof of the cube only (see Figure 26). A negligible difference was found for the
windward and leeward face of the cube. Further refinements made to the grid which
involved increasing the number of points around the cube from 40
3
to 60
3
(see Figure
27) showed no further changes to the predicted pressures on the roof of the cube. As it
was not possible to obtain sufficiently converged results for the DSM only the RNG
and non-linear k- models were run to the highest levels of grid refinement.
Figures 28 and 29 show the roof pressures distributions for grid refinement test 2 with
the cube at 45 to the incident wind, for the non-linear k- model. These tests used
four different grids incorporating 150,000 to 1,250000 nodes. As a negligible
difference was found for the windward and leeward face of the cube the main aim of
83
these tests was to attempt to adequately resolve the delta wing vortex. The grid
refinements incorporated 20
3
, 40
3
, 60
3
and 90
3
grid points around the actual cube. The
graphs show that as the number of grid points increased the simulations were able to
more accurately predict the sharp negative peak at the front corner. Unfortunately it
was not possible to obtain a grid independent solution for the flow at the very front
corner of the skewed cube, due to computer hardware memory limits. This suggested
that the vortices produced have a very high angular velocity and therefore are very
small and difficult to resolve. Figure 30 highlights the predicted increase in strength
of the delta wing vortex (middle section of the cube), with refinement of the grid. Due
to the excessive run times of these simulations it was decided not to test the standard
and MMK k- models, which were found to produce poor results for relatively coarse
meshes, as described later. Furthermore, it was not possible to obtain adequately
converged solutions when using the differential stress model
Figure 31 shows grid refinement tests for the SSB, with sharp eaves at a normal
orientation, for comparisons between 220,000 and 440,000 and 750,000 node model
runs. In a similar manner to the earlier tests there was a negligible difference between
the pressures for the front and leeward faces of the building. It can be seen that, for
the windward roof, there is a negligible difference between the latter two grids
consequently the majority of SSB simulations were undertaken with approximately
500,000 nodes. These results were initially surprising when considering the
refinements needed for the 6m cube. Analysis of the results for the 6m cube showed
that only the roof area produced changes with successive grid refinements. Therefore
it appears that as the roof of the SSB is pitched the predicted velocity gradients on the
roof of the SSB are less severe therefore reducing the level of grid resolution required.
This is particularly so for the SSB with curved eaves where there is no flow
separation.
Grid refinement tests for the experimental wall structure, from the original model with
550,000 nodes to approximately 1 million nodes, showed no change in the predicted
mean pressure coefficients (see Figure 32).
6.3.10 Double Precision Checks
All the numerical simulations for this project were run using single precision
numerical accuracy. The effect of rounding off the solutions over a number of
iterations can be a source of error in the calculations. Therefore a small number of
84
tests were undertaken to assess any possible effects of running the simulations with
single precision accuracy. These tests used the RNG k- turbulence model and CCCT
convective differencing and the results showed that there was no difference between
the single precision and double numerical precision results. Residual convergence
histories can be seen on Figures 33 and 34, for a 6m cube simulation.
6.3.11 6m Cube Model - Reynolds Number Effects, changes to the Inlet Velocity
A number of simulations were run using the non-linear k- model and CCCT
differencing with inlet velocities of 0.1 and 10 m/s, set at a reference height of 10
metres, in order to assess the effects of increasing the Reynolds number by a factor of
100.
Figure 35 shows the normalised roof pressure distribution for the two simulations
with differing velocities. It is clear from this graph that the change in Reynolds
number from approximately 600,000 to 60 million had relatively little effect.
Referring to Figure 36, it can be seen that there is an increase in the roof vortex
recirculation strength of approximately 6 percent.
The latter results are not a surprise as one would expect a larger roof recirculation
vortex as the Reynolds number increases due to a reduction in the effects of the fluid
viscosity, i.e. Reynolds number vortex effects (Hoxey 1998). Consequently although
only minor changes to the model predictions are shown all tests were run at a standard
wind speed of 10 m/s at cube height. This velocity compared with the full-scale
experimental wind speed data collection range of between 6 and 10 m/s. The small
differences in wind speed between the full-scale and CFD data should have a
negligible effect on the comparisons.
6.3.12 6m Cube Model - Effects of the Differencing Scheme.
All the simulations undertaken used both the CCCT (section 3.5.5) and Hybrid
(section 3.5.3) differencing schemes. These particular schemes were chosen as they
represent the most popular stable high and low order accuracy differencing schemes.
Attempts to use the QUICK scheme with the k- model were unsuccessful, due to the
over and undershoots which are common with this scheme because of its lack of the
boundedness property (section 3.5.4). Tests using upwind differencing were not
85
considered as it is very similar to the hybrid scheme when used in high Reynolds
number flows (see section 3.5.3).
Figures 37 and 38 show the simulation results for the pressure coefficient and
turbulent kinetic energy values calculated on the roof of the cube for both CCCT and
hybrid differencing. Figure 39 shows the roof recirculation velocity calculated for the
two differencing schemes.
The effects of numerical viscosity were apparent with all model simulations with the
first order accurate hybrid differencing scheme. The most notable effects were an
increase in the windward face and roof front edge mean pressure coefficients brought
about by a general increase in the effective fluid viscosity and therefore an increase in
the turbulent kinetic energy levels. For the 6m cube normal to the incident wind, with
both the RNG and non-linear k- model simulations, the extra numerical viscosity
resulted in a 5-7 percent reduction in the length of the roof vortex.
The hybrid scheme, utilising upwind differencing is showing the effects of false
diffusion, apparent in multi-dimensional flows if the velocity vector is not parallel to
one of the coordinate directions. This can be graphically demonstrated by the
calculation of a scalar property in a domain where the flow is at an angle of 45 to the
Cartesian grid. By setting the flow to be laminar and using symmetry boundary
conditions there should be no turbulent and minimal viscous diffusion of the scalar.
The effects of false diffusion can be clearly seen on Figure 40 causing the scalar to
disperse with distance along the domain for the hybrid scheme. Figure 41 clearly
shows the reduced diffusion predicted by the higher order CCCT differencing scheme.
It is therefore recommended that all CFD simulations should incorporate higher order
convection differencing to achieve the highest levels of accuracy.
6.3.13 Use of High Reynolds Turbulence Models
For the tests undertaken in this project it has only been possible to use turbulence
models that use wall functions in the near wall region. There are errors in using the
oversimplifications for the wall functions in wind engineering simulations as reported
by Murakami (1990). Unfortunately due to the size of the computational domains and
the significant inlet velocities and high Reynolds numbers (see section 6.3.4) it proved
impossible to mesh the geometries to allow sufficient resolution of the viscous sub-
layer for the low Reynolds models. Low Reynolds number turbulence models require
86
approximately 30 nodes in the viscous sub layer alone with the two layer models
requiring between 15 and 20 nodes. Therefore it is conservatively estimated that 1.5
to 2 million nodes would be required for the 6m cube model with the larger SSB
geometries requiring an even greater number. An indication of the increased
computational overheads required by the use of these models is given by Lakehal and
Rodi (1997). It is stated that experiments using the standard k- model with wall
functions took 15 minutes to complete while similar two layer and low Reynolds
number model tests required 6 and 20 hours respectively. Furthermore, it is reported
by Hanjalic et al (1998) that two layer and low Reynolds number models produced
only slight improvements to the predicted flow field around a surface mounted cube.
This indicates that wind engineering flows are dominated by convection and diffusion
of the large scale vortices and eddies that form around the cube and less by the small
scale structures in the flow.
Finally it has been reported that the use of low Reynolds number turbulence models
may significantly reduce the numerical stability of the model. Menter and Grotjen
(1999) state that this is due to the highly non-linear damping terms used on the
transport equations. Nonetheless, low Reynolds number turbulence models have been
used to provide comparative data for the large eddy simulation runs reported in
chapter 7, which use a revised computational domain to aid resolution of the sub-
layer.
6.3.14 Solution Convergence
All the tests undertaken were ran to achieve high levels of convergence. For the eddy
viscosity based models the minimal residual reduction factors achieved (that is the
ratio of the second to last residual error values) were 1 10
5
. Meeting this criteria for
the differential stress model proved more of a challenge, particularly so when using
CCCT differencing, therefore for the DSM the minimal residual reduction factors
achieved were 1 10
3
to 1 10
5
. If there were difficulties in obtaining residual
reduction factors of greater than 1 10
3
further checks were carried out, which
included analysis of the flow field for flow symmetry, to ensure adequate solution
convergence had been attained.
87
6.3.15 Calculation of Pressure Coefficients.
The standard method of detailing pressure distributions in wind engineering are wind
induced dynamic pressure coefficients, C
p
, where ambient static pressures cancel out,
for full details see Cook (1986b). Referring to Patankar (1980):
If the absolute value of pressure were used, then round off errors arise in
calculating differences for pressure at each face of the control volume. It is
therefore best to set p = 0 as a reference pressure at a suitable grid point and to
calculate all other values of p as pressure relative to this reference value.
(Patankar 1980, p131)
Therefore the CFX derived pressures are in an appropriate form to allow correct
calculation of the pressure coefficients which are calculated by dividing the CFD
derived pressure by the dynamic pressure calculated at 6m, obtained from
2
2
1
V .
Nonetheless there is in fact a small error associated with this assumption as used in
the following work. There are extra contributions to the dynamic pressure that need to
be accounted for due to the effects of the turbulent kinetic energy in the flow. If the
free stream turbulent kinetic energy levels are relatively low, at the reference height,
then the additional terms will be small and of little consequence. The following text,
provided by Dr Hoxey (1999) of the SRI, derives this additional contribution for the
CFD calculations.
2
2
1
u q (6.19)
2
2
1
u q (6.20)
( )
2
'
2
1
u u + (6.21)
( )
2 2
' ' 2
2
1
u u u u + + (6.22)
0 ' since '
2
1
2 2
+ u u u (6.23)
88
or

,
_

+
2
2
2
'
1
2
1
u
u
u q (6.24)
) 1 (
2
1
2 2
I u + (6.25)
where I = turbulence intensity and in a complete analysis should include u, v and w.
Therefore Equation 6.25 becomes:
) (
2
1
2 2 2
w v u q + + (6.26)
hence
2 2
2 2 2
2
2 ' ' '
u
TKE
u
w v u
I

+ +
(6.27)
where TKE is the free stream turbulent kinetic energy at the reference height, which
for the cube model is equal to 1.187m
2
/s
2
. Therefore the actual value of dynamic
pressure at the reference height of 6m for the cube, due to the mean and turbulent
contributions, is:
)
) 10 (
187 . 1 2
1 ( ) 10 (
2
1
2
2

+ q (6.28)
) 024 . 1 ( ) 10 (
2
1
2
(6.29)
The effect of the turbulent kinetic energy in the flow will result in an increase of the
dynamic pressure of approximately 2.4 percent. Calculations for the SSB and
experimental wall structure provided a similar result. Therefore it is apparent that the
additional effects of the turbulent kinetic energy in the CFD calculations for the
derivation of dynamic pressure are relatively small. The effects for the full-scale
experimental work are more profound with an increase in the dynamic pressure of
approximately 8 percent (Hoxey 1999).
6.3.16 Accuracy test summary.
A summary of the accuracy test results is given in Table 4 below. This table shows
the sensitivity of the models tested to the changes discussed above as a worst case
percentage error. Therefore the values given represent the maximum worst case
89
change in the predictions at some location on the front or roof of the model. This
method of assessing the accuracy of the turbulence model with remain consistent
throughout chapter 6.
Percentage error. Test information.
Windward face Roof
Domain size Negligible Negligible
Revision to the wall functions 5 5
Ground Roughness 0.01 to 0.04m 15 15
Building Roughness 0.005 to 0.001m Negligible Negligible
y
+
checks ----- -----
Grid refinement cube normal Negligible 10-15
Grid refinement cube skewed Negligible Unknown
Grid refinement experimental wall Negligible Negligible
Grid refinement SSB 2 12
Double precision Negligible Negligible
Reynolds number effects Pressure 0 Pressure 0 / velocity 6
Full-scale press. coeff. error (approx.) t 1-2 t 1-2
Calc. of dynamic pressure coefficient. 2.4 2.4
Changes to the differencing scheme Pressure 10 Pressure 12 / velocity 30
Table 4: Accuracy test summary.
6.4 Turbulence Model Validation The 2-Dimensional Backward Facing
Step
As this project required the testing of a number of new turbulence models, as detailed
in chapter 5, it was important to prove that the models were correctly coded and that
they were capable of providing acceptable or improved flow field predictions.
Since the 1980 AFOSRHTTM Stanford conference on complex turbulent flows the
two-dimensional backward facing step has been established as a standard method of
validating turbulence model performance and accuracy. Therefore this provided the
ideal test case for the new turbulence models.
Basora and Younis (1992) have previously compared standard k- model results with
wind tunnel based 2D back-step tests undertaken by Kim et al (1980). Therefore to
effectively calibrate a 2D back-step model for use as a test case, it was necessary to
match the standard k- model results produced by CFX4.2 to those found by Basora
and Younis (1992). It was then possible to test the revised turbulence models and
directly compare their accuracy to the experimental data.
90
Although the above mentioned models required access to source code to allow the
author to implement them into CFX4.2 it was in fact possible to implement the MMK
k- model through the use of the user defined FORTRAN routines. As such this
model was programmed and tested by Dr Andrew Quinn, of the Silsoe Research
Institute, at an early stage in this project and kindly passed on to allow full testing in
the following chapters.
Table 5 shows the results.
Turbulence model Reattachment. length
Std k- 5.4h
Non-linear quadratic (Speziale) ------
Non-linear quadratic (Craft et al 1996) 6.5h
Non-linear cubic (Craft et al 1996) 7.15h
D.S.M (no wall reflection) 6.5h
D.S.M (wall reflection terms) ------
Experimental (Kim et al 1980) 7.0h
Table 5: Reattachment lengths for backward facing step
(h = step height)
The Speziale (1987) quadratic non-linear turbulence model was tested for this case
but could not provide converged solutions. It has been reported by Speziale and Ngo
(1988) that there are general convergence problems due to high gradients returned by
the second derivatives of velocity brought about by the use of the Oldroyd derivative
term (see Equation 5.16). As such this model will not be tested further in this report.
Preliminary testing of the DSM with the Gibson and Launder (1978) wall reflection
terms for the 2D back-step test case was unsuccessful. The addition of these terms as
sources to the Reynolds stress transport equation resulted in extremely poor numerical
stability and a resulting lack of convergence.
Comparisons between the standard k- model and the quadratic non-linear k- model,
detailed by Craft et al (1996), show no measurable difference in solution times. The
extra time required per iteration of the non-linear model was offset by an increase in
the speed of convergence, as this model avoids the unrealisable values returned by the
production of the turbulent kinetic energy term in the standard k- model.
In terms of the accuracy of the models tested it can be seen that the standard k-
model predicts the worst results with an error of approximately 23 percent. This was
91
followed by the DSM and the quadratic and cubic non-linear k- models with errors of
7 and 2 percent respectively. These initial results combined with extensive checks of
the new code added to CFX effectively validate the revised models and gave an early
indication of possible improvements in the predicted flow fields.
6.5 CFD and Full-Scale Test Results
6.5.1 Cube Normal to the Incident Wind - Introduction
All of these tests included the 6m cube at a normal orientation to the incident wind
with the computational domain constructed, as detailed in Figure 42. Further
representations of the computational domain and mesh can be seen in Figures 43 and
44 respectively.
A total of 8 turbulence models were tested as discussed in section 6.1. The Speziale
(1987) version of the quadratic model and the Craft et al (1996) version of the cubic
non-linear k- model were omitted due to their inability to meet the convergence
criteria adopted. Discussions with Dickens (1999) of Computational Dynamics
suggests that on the basis of his experience, of incorporating these models into Star-
CD, the numerical instabilities are simply a result of the numerous extra terms present
in the cubic model. Further refinements may be possible to improve the overall
stability of this model.
Both the two layer model and the DSM with wall reflection terms were omitted from
these tests due to the fact that a number of wind engineering simulations have already
been undertaken and reported on (detailed in chapter 5). Furthermore, as explained
previously, the need to tailor the DSM with wall reflection terms to each individual
test case immediately made it difficult to justify as a generic turbulence model to be
used by the wind engineering community.
All graphs and data presented are for cube centreline measurement positions unless
otherwise stated, as shown in Figure 45.
Although it would have been preferable to directly compare mean turbulent kinetic
energy and Reynolds stress profiles from the experiments undertaken at the SRI with
CFD data this was, unfortunately, not possible. At present only the mean pressure
92
coefficients and recirculation vortex reattachment length data are available.
Nonetheless, this information was sufficient to clearly distinguish which CFD models
were the most accurate in wind engineering applications.
6.5.2 Roof / Wake Flow Field Patterns
Turbulence Model. Roof Reattachment. Wake Reattachment.
Standard k - epsilon. No separat i on. 2. 1H
M. M. K. k - epsi l on. No reat t achment . 3.12H
RNG k - epsi l on. 0. 84H 2. 5H
Non-lin. Quad. , k-eps. 0. 75H 2. 2H
Differential stress. No reat t achment . 2. 1H
Exp. ((Hoxey 1999) 0. 5-0. 6H 1. 2-1. 4H
Table 6 : Reattachment lengths: cube normal to the incident wind.
( H = cube dimension)
Table 6 shows the CFD and experimentally obtained reattachment locations for the
roof and wake recirculation vortices. The non-linear and RNG k- models predicted
the most accurate roof reattachment lengths with respective errors of between 25 to 50
percent and 40 to 68 percent. The standard k- model predicted no separation in
contrast to both the MMK k- and differential stress models, which did not predict
any reattachment of the roof vortex. For the wake recirculation the differential stress
and standard k- model simulated the most accurate vortex followed by the quadratic
non-linear k- model. Surprisingly the standard k- model predicted a relatively
accurate wake recirculation length and the MMK poorly calculated the wake flow
field. The results for standard k- models wake were also found by Hanjalic et al
(1998). It should be noted that none of the models tested was able to accurately
predict the length of the wake vortex. The comparative vortex strengths are shown on
Figures 46 and 47. Figures 48 to 52 show the roof velocity vectors for the turbulence
models tested.
6.5.3 Normal Cube Mean Pressure Distributions
Figure 53 shows the centreline windward face mean pressure coefficient distributions.
It can be clearly seen that the standard k- model over predicted the pressure
distribution by approximately 25 to 30 percent at the flow stagnation point. The
remaining four models all predicted relatively similar results within approximately 10
percent of the experimentally obtained values. The differential stress model calculated
93
the most accurate result with a pressure distribution marginally closer to the
experimental value. It is also apparent from the results that the CFD predicted
stagnation points are approximately 0.5m higher than that obtained by experiment.
Figure 54 shows the centreline pressure coefficient distributions for the roof of the
cube. The differential stress and the non-linear quadratic model predicted similar
pressure distributions and the most accurate suction peak pressures within 5 percent of
the peak mean experimental values. The RNG and MMK k- models slightly over and
under predicted the peak pressure at the front edge of the cube respectively, although
an accurate assessment was difficult due to the fact that the nearest experimental point
was 0.4m from the edge of the cube. The standard k- model over estimated the peak
suction pressure by approximately 65 percent. It is apparent that none of the models
tested could accurately predict the pressure distribution over the remainder of the roof
with even the best CFD results calculating errors of 20 to 30 percent. Grid refinement
tests for the RNG and non-linear k- models resulted in an overall improvement of 5
to 10 percent for the roof pressure distribution.
It should be noted that there are a few inconsistent experimental points both on the
side wall and most noticeably on the roof of the cube that cannot be fully explained.
According to Hoxey et al (1999) these errors are most likely due to approach flow
turbulence intensity variations. These possible errors are to be the subject of further
investigations by the SRI (Hoxey et al 1999).
The CFD results presented thus far are not as good as might be hoped, although they
do display the correct magnitude and trends in many cases.
The leeward pressure distributions can be seen on Figure 55. It was again apparent
that none of the models tested could accurately predict the experimentally obtained
distribution. The best results were obtained from the quadratic non-linear k- model,
although errors of approximately 30 to 35 percent were still apparent. The next most
accurate model was the RNG k- model. The DSM predicted the worst results with
errors of up to 50 percent. It appears that the under predictions of negative pressure
are a consequence of the over prediction of the wake recirculation and the
corresponding lack of velocity deficit. These results confirm the need to accurately
simulate the flow field around the bluff body. No improvement was found with grid
refinement and there is some question over the middle experimental value obtained
which appears to be out of sequence with the remaining values.
94
The side face and lateral roof pressure distribution can be seen on Figures 56 and 57
respectively. The experimental results for the side face showed a slow increase in
negative pressure as the wind velocity increases with height. It can be seen that this
trend was well reproduced by the MMK and RNG k- models but not so well by the
remaining models. The standard k- model predicted the highest negative pressures
for the first 3m height of the side face, possibly for the same reason that it failed to
calculate flow recirculation over the roof of the cube, due to excessive prediction of
turbulent kinetic energy as described below. Between 3 and 6m in height it is apparent
that the non-linear model predicted the most accurate values with errors in the side
and roof pressure prediction ranging from 20 to 30 percent and 30 to 40 percent
respectively. Figure 57 further demonstrates the results shown by Figure 54,
highlighting the inability of the models tested to accurately calculate the pressure
distribution over the roof of the cube. The trend shown by the experiment represents
the reduction in size and strength of the roof recirculation zone as the sides of the
cube are approached and the mass of air flowing over the roof is reduced. It is clear
that the experimental results are affected to a much larger extent by this three
dimensional phenomenon, due to the significant difference between the centre and
edge negative pressures. This perhaps suggests that the roof recirculation zone in the
full-scale experiments contains a much stronger vortex than that predicted by CFD.
This would further explain the slow reduction in negative pressure, for the
experimental results, as the roof is traversed as shown on Figure number 54.
6.5.4 Turbulent Kinetic Energy Distributions.
Figures 58 and 59, show that at the front corner of the cube the standard k- model
predicted levels of turbulent kinetic energy that are several factors higher than those
predicted by the differential stress model. The MMK k- model predicted the lowest
values of turbulent kinetic energy closely followed by the differential stress model.
The non-linear and RNG k- models calculated the next highest turbulent kinetic
energy levels. In addition, Figure 59 shows that, both these models predicted a slight
increase in turbulent kinetic energy in the region of the flow recirculation. Full kinetic
energy contour plots for the range of models tested can be seen in Figures 60 to 64.
The excessive stagnation and front corner turbulent kinetic energy predictions by the
standard k- model can be clearly seen, which are in contrast to the predictions by the
other models. Full discussion of these results and their effects is given in section 6.9.
95
6.5.5 Pressure and Velocity Distributions Cube orientated at 45
The line positions shown (see Figure 65) indicate the locations used when recording
data with the CFX-Visualise linegraph package. These cover the locations of the
pressure taps for the full-scale experiments. All other experimental data is as
described in the previous sections.
Figure 66 shows the CFD predicted strength of the lateral velocity over the roof of the
cube. This graph therefore shows the relative strength of the delta wing vortices
predicted. The differential stress model and the non-linear and RNG k- models all
predicted the highest velocities closely followed by the MMK k- model. The
standard k- model, as would be expected, predicted the weakest vortices. Figure 67
shows a vector plot of the counter rotating vortices at the same measurement location
used for the data shown on Figure 66, at the middle section of the cube.
Figure 68 shows the CFD and experimentally obtained pressure coefficient
distribution for line number 1 (Figure 65). The differential stress model appears to
calculate the most accurate values with the predicted pressure distribution within 15
percent of those obtained from experiment. The standard k- model again appears to
over predict the mean pressure at the stagnation point, although with a lower error
than for the normal cube due to the reduced flow impingement. The remaining models
all calculated similar front face pressures with an error of approximately 25 percent.
In addition, all the CFD models simulated a flow stagnation point approximately 1.0
to 1.5m metres higher than that obtained from the full-scale experiment.
Figure 69 shows the roof pressure distributions for measurement line 2. The
differential stress model predicted the highest suction pressure, perhaps as a
consequence of calculating the strongest delta wing vortices. In this case the standard
k- model calculated the least negative pressures due to the prediction of very weak
delta wing vortices. Successive refinements of the grid over the cube achieved a 15
percent increase in the peak pressure over the roof and improvements to the
distribution of the pressures, see Figure 70. Unfortunately computational limitations
did not allow calculation of a grid independent flow field for the roof. Nevertheless
the overall predicted pressure distribution is relatively accurate for the turbulence
models tested with an increase in accuracy of both the peak pressure and remaining
profile upon successive grid refinements.
Referring to Figure 71, which shows the predicted pressure distributions for
measurement line number 3, it can be seen that non-linear model calculated the most
96
accurate values with an error of approximately 40 percent. In addition, this model
calculated a relatively accurate pressure distribution profile with a small increase at
the centre of the cube. The DSM closely followed the non-linear model with the
standard k- model predicting the worst distribution with errors of up to 60 percent.
None of the latter models correctly predicted the profile of mean pressure on the
leeward face of the cube.
6.6 The Silsoe Structures Building (SSB)
6.6.1 Introduction
A number of different SSB geometries and orientations have been tested, for
comparison to full-scale data, including the following:
1. SSB with sharp eaves normal to the flow
2. SSB with curved eaves at both normal and transverse orientations to the flow.
All the considerations detailed in the previous sections, relating to the obtaining of
high quality CFD results, were fully applied to the SSB test case. Furthermore,
considerations beyond those highlighted for the cube model were necessary due to the
further complexity of this structure. Bickerton et al (1996) has reported that special
care has to be taken to ensure that the method of domain construction and direction of
the grid lines does not influence the flow field. This is particularly important in the
region of the leeward roof where grid lines parallel to the roof may, for shallow
angles, influence whether the flow remains attached or separated. Consequently tests
were undertaken with different block structures that fully addressed these points and
showed that for this case there was no noticeable effect to the roof flow field or
pressures distribution calculated.
Due to the relatively poor results found from the application of the MMK k- model
to the 6m cube (section 6.6) it was decided not to test the model further.
Figure 72 details the SSB geometry and pressure measurement locations that extend
along the position shown from the windward to the leeward face:
97
Full details of the experimental procedure, the Silsoe structures building (SSB) and
further references are given in section 6.2.2.
6.6.2 Flow Field and Pressure Coefficient Distribution SSB with Sharp Eaves
at a Normal Orientation to the Incident Wind
For this orientation all CFD simulations, bar the DSM, were run with a reference
velocity of 10 m/s at the ridge height. In order to obtain converged results with the
DSM it was necessary to reduce the velocity to 1 m/s. Previous tests (see section
6.4.1) with the cube model and the SSB have shown a minimal difference in the
normalised predicted flow field and pressure distributions for the differing velocities.
This is due to the fact that both velocities involved sufficiently high Reynolds
numbers.
Turbulence model Reattachment length
RNG k- 1.8m
Std k- No separation
Non-linear quadratic k- 1.2m
DSM 2.0m
Experimental (Hoxey 1999) 1.0 1.5m
Table 7: Reattachment lengths for the front corner vortex: the SSB
Table 7 provides reattachment length detail for the roof vortex formed on the
windward roof as a consequence of flow separation at the sharp front edge of the
building. It can be seen that the non-linear model predicted the most accurate vortex
length of all the models tested. It is also apparent that the differential stress model
predicted flow reattachment in this case as a consequence of the pitched roof.
Figure 73 details the windward face pressure distribution at measurement line 3 (see
Figure 72). It can be seen that all four CFD predictions follow the correct trend with
the non-linear k- model predicting the most accurate pressure distribution closely
followed by the RNG model with respective worst case errors of approximately 20
and 35 percent. In this case the differential stress model shows the worst prediction
with errors of over 70 percent. Referring to the discussion offered for the 6m cube,
this is a rather surprising result, as it can be seen that the differential stress model
predicts a much lower level of turbulent kinetic energy than the standard k- model
(see Figure 74). For this test case it appears that the non-linear and standard k-
98
models predicted the correct flow stagnation point location with the RNG and
differential stress models under predicting the height by approximately 0.5m.
Figure 75 shows the windward roof centreline pressure distribution at measurement
line number 3. It can be seen that the differential stress and non-linear models
predicted relatively accurate and similar maximum mean peak pressures at the front
edge of the building, although it should be stated that the nearest experimental value
was approximately 0.5 metres from the windward edge of the SSB. For the remaining
profile of the pressure distribution it appears that the differential stress model
predicted a marginally more accurate distribution with a shallower gradient as the
pressure recovers. The RNG and standard k- models calculated a relatively similar
peak suction pressure, significantly higher than predicted by experiment, with the
standard model predicting a much steeper pressure recovery. The high negative
pressure predicted by these models was due to the excessive calculation of turbulent
kinetic energy at the windward face and roof of the building (see Figures 74 and 76).
The steep gradient of the pressure recovery calculated by the standard k- model is
due to the lack of a predicted separation zone at the front edge as simulated by the
remaining models and present in the experimentally obtained data. For the overall
pressure distributions the percentage error in the CFD calculations ranged from 20 to
40 percent. These results effectively mirror those produced for the 6m cube whereby
the CFD models all calculated too rapid a pressure recovery along the roof. Figure 77
shows kinetic energy contours for the full SSB simulation, using the non-linear k-
model. The increased values of turbulent kinetic energy can clearly be seen at the roof
recirculation point.
Figure 78 shows the leeward roof pressure distribution at measurement line number 3.
It was apparent that the non-linear and standard k- models showed very good
agreement with the pressures obtained by experiment. The RNG k- model showed an
under prediction of the pressure of approximately 5 percent, with the differential
stress model calculating the least accurate pressures, with an error of approximately
20 percent near the ridge of the building. The flow over the latter half of the
windward roof and over the whole of the leeward roof remained attached perhaps
explaining the improved accuracy of the results for the different turbulence models
tested. Due to the minimal number of experimental pressure measurement points it
was difficult to interpolate the results and ascertain whether the CFD derived peak
pressures at the ridge of the building were accurate or not.
99
For the leeward face pressure distribution at measurement line number 3 (see Figure
79) it was interesting to find that the standard k- model predicted the highest negative
and most accurate pressure distribution. This was most likely due to the velocity
deficit in the wake due to the excessive prediction of eddy viscosity in the simulation.
This effect resulted in a relatively accurate wake vortex prediction for the 6m cube
model. The differential stress and non-linear models predicted similar distributions
with errors of around 15 percent. As noted earlier, the lack of experimental data in
this region made an accurate assessment of the model performance particularly
difficult.
6.6.3 Pressure Coefficient Distribution: The SSB with Curved Eaves at a
Normal Orientation to the Incident Wind.
For this orientation all the models tested were ran with an inlet velocity of 10m/s at a
reference height of 5.3m (ridge height).
Figure 80 details the windward face pressure distribution for the SSB, with curved
eaves, at measurement line number 3. Interestingly this graph shows that CFD under
predicted the pressures for all the turbulence models tested. This is in contrast to
results discussed earlier for the SSB with sharp eaves where the CFD calculated
pressures were all over predicted. The experimental values for the two cases are
relatively similar with a slightly lower pressure coefficient distribution for the curved
eaves case. This was most likely due to the smaller surface area presented at the front
face due to the curved eaves. For both tests the DSM predicted the highest front face
pressure distribution which, for this test, proved to be the most accurate with errors
ranging from 5 to 25 percent. The non-linear model, which on the SSB with sharp
eaves calculated the most accurate and lowest pressure distribution, in fact calculated
the worst pressure distribution for this case with errors of approximately 35 percent
for the majority of the front face. The RNG model predicted similar results to the non-
linear model. All the models tested predicted the correct trends for the pressure
distribution.
Figure 81 shows the windward face roof pressure distribution for measurement line
number 3. It was apparent that the non-linear model predicted the most accurate
distribution near the windward corner of the roof followed by the DSM. After
approximately half way along the roof the standard k- model predicted the best
pressure distribution followed by the non-linear model with the least accurate
100
prediction by the DSM. For this case there is no prediction of a separation vortex due
to the curved eaves, consequently the problems encountered by the standard k-
model for the 6m cube simulations are not apparent. Furthermore, it is again clear that
CFD was able to provide far more accurate results for a test case where the flow
remained attached, such as this case for the windward roof of the SSB with curved
eaves.
Figure 82 shows the pressure distribution for the leeward roof at measurement line
number 3. This graph shows that for the first metre along the measurement line the
standard k- model actually predicted the most accurate pressure distribution. This
was perhaps a result of the standard k- models relatively accurate prediction of the
windward roof pressure distribution (latter half). This was combined with the
prediction of a high level of turbulent kinetic energy, as shown on Figure 83, resulting
in higher suction pressures at the ridge. The DSM was the next most accurate model
along this section. Beyond the front section of the leeward roof the remaining
turbulence models predicted very similar pressure distributions which accurately
followed the experimental values with the standard k- model over predicting the
pressure distribution.
Figure 84 shows the leeward face pressure distributions. All the models tested
followed the same pattern, as displayed for all the previous simulations, in under
predicting the distributions. For this test the non-linear k- model predicted the most
accurate results followed by the RNG model with the standard k- model and DSM
predicting the least accurate distributions.
6.6.4 Pressure Coefficient Distribution SSB Curved Eaves / Transverse
Orientation
For this set of tests measurements were taken with the Silsoe structures building in a
transverse orientation to the incident wind, as shown in Figure 72. Three measurement
positions were chosen, as shown, to determine the pressure distributions along the
building.
In this case it was found that the DSM would not produce results that met the
convergence criteria. This problem continued after numerous attempts to improve the
model by re-meshing the domain with lower geometric progression ratios, restarts
from converged k- model solutions, reductions in the inlet velocity and changes to
101
the solution parameters in the command language. Consequently these tests were
completed using three k- model variants, the standard, RNG and non-linear models.
Figures 85 and 86 show the CFD and experimentally obtained pressure distributions
for measurement line number one on the side face and roof of the building
respectively. Figure 85 shows that the non-linear and standard k- model produced the
most accurate pressure coefficients with errors ranging from 25 to 40 percent for the
front face of the building. Interestingly the non-linear model showed a profile that was
similar to that predicted for the side face of the 6m cube. Initially there is an increase
in the negative pressure coefficient near the ground followed by a decrease after
which all the models follow the same pattern. At present it is unclear why this
particular profile is occurring. The general profiles calculated by the RNG and
standard k- models more accurately followed the trend shown by the experimental
results.
Figure 86 shows that the non-linear model produced pressure coefficients within 10
percent of the experimentally obtained values. The RNG model is the next most
accurate model with a consistent error of approximately 30 percent followed by the
standard k- model with modelling errors of up to 65 percent near the curved eaves.
For this case the standard k- model incorrectly predicted the profile along the roof of
the SSB. This is quite possibly due to the failure of the model to calculate a
recirculation vortex along the roof as detailed, for the non-linear k- model
simulation, in Figure 87. A recirculation zone occurs which is strongest at the ridge
and non-existent at the edge of the roof. It appears that the standard k- model
predicted that the flow remains attached to the roof for the majority of the
measurement line.
Figure 88 shows the side face pressure coefficient distributions for measurement line
number 2 (Figure 72). This graph shows that the RNG k- model predicted the most
accurate results followed by the non-linear model and standard k- model which over
and under predicted the negative pressures, respectively. The roof pressure
coefficients for measurement line number 2 are shown on Figure 89. This shows that
the non-linear and RNG models calculated the most accurate peak pressures but failed
to predict the correct gradient of the line as the pressure recovers along the roof. The
standard k- model predicted a much lower negative pressure distribution than the
other models due to the early recovery of pressure along the roof caused by the lack of
simulation of a roof recirculation zone, as previously detailed for the 6m cube tests.
102
Figures 90 and 91 show the side face and roof pressure coefficients for measurement
line number 3 respectively. Both graphs show how all the models tested significantly
over predicted the suction pressure distributions, with errors up to approximately 100
percent, on comparison to the experimentally obtained results. The general over
prediction of pressures by CFD at measurement line number 3 may be due to an over
prediction of the roof separation vortex as shown in Figure 87. It is possible, although
by no means proven, that the vortex had reattached in the experiment leading to a
faster pressure recovery than that predicted by the CFD tests. Overall, the transverse
SSB test case has highlighted significant CFD modelling errors.
6.7 The Silsoe Experimental Wall Structure
6.7.1 Introduction
A schematic of the experimental wall, including information on measurement
locations, is given in Figure 92.
Full details and further references for the experimental wall were given in section
6.2.3. In order to obtain high quality CFD data these tests were subject to the same
accuracy checks as the 6m cube model and the SSB. All tests used 1:1 scale CFD
models and the atmospheric boundary layer inlet conditions discussed in section 6.3.5.
Due to the previous convergence problems found when using the differential stress
model it was decided to compare the experimental wall data to CFD runs of the
standard, RNG and non-linear k- models only.
6.7.2 Pressure Coefficient Distribution: Experimental Wall Windward and
Leeward Faces for a 0 (normal) Incident Wind.
Figures 93 and 94 show the CFD and experimentally obtained windward and leeward
centre face pressure coefficient distributions for the experimental wall setup shown in
Figures 16 and 92. For the windward face it can be seen that the non-linear and
standard k- models predicted relatively accurate pressure distributions at the centre
section of the wall. The RNG k- model over predicted the pressure coefficients at the
centre of the wall with an approximate error of 20 percent. All the CFD models tested
predicted a large pressure differential between the centre and edges of the wall, in
contrast to the experimental data which showed relatively little change in the pressure
103
coefficients along the wall. The experimental results are therefore surprising
considering that the flow is subjected to a greater obstruction at the centre than
towards the edges, thus a higher stagnation pressure should be experienced at the wall
centre. Referring to Figure 94 for the leeward face it can be seen that the standard k-
model calculated the most accurate results at the centre of the wall and least accurate
at the edges with an error of approximately 80 percent. The non-linear model
predicted the most accurate results overall with an average error of between 10 and 15
percent. The RNG k- model calculated an error at the wall centre of approximately
40 percent with relatively accurate wall edge pressure predictions. This model
appeared to more accurately follow the pattern shown by the experimental results
although unfortunately it significantly over exaggerated the general trend towards a
decrease in suction pressure at the wall centre.
6.7.3 Pressure Coefficient Distribution: Experimental Wall Windward and
Leeward Faces for a 30 Incident Wind.
Figures 95 and 96 show the windward and leeward face mean pressure coefficient
distributions for the experimental wall structure skewed at 30 to the incident wind.
Referring to Figure 95 it can be seen that all the models tested over predicted the
mean windward face pressure distribution with maximum errors of approximately 75
percent for the standard and RNG k- models. The non-linear k- model provided
improved results with errors of up to 40 percent. The CFD calculated leeward face
pressures were generally much closer to the values obtained from experiment with the
RNG and non-linear k- models both calculating accurate pressure distributions along
the length of the wall. Unfortunately these models were unable to predict the
maximum peak negative pressure at the leading edge of the wall. The standard k-
model predicted a peak negative pressure approximately 25 percent higher than that
obtained from the experiment. Referring to Figures 97 and 98, it can be seen that the
standard model calculated a higher level of turbulent kinetic energy in the wake
region and at the leading edge of the wall, than the non-linear model. This had the
effect of reducing the strength of the recirculation vortex on the leeward side of the
leading edge of the wall (see Figures 99 and 100) resulting in a much higher peak
negative pressure at this location. The reason for this was similar to that described for
the 6m cube model tests (section 6.6.4)
104
6.7.4 Pressure Coefficient Distribut ion: Experimental Wall Windward and
Leeward Faces for a 60 Incident Wind.
Figures 101 and 102 show the experimentally and computationally obtained
windward and leeward face mean pressure coefficient distributions for the
experimental wall structure skewed at 60 to the incident wind. These results showed
a general reduction in the percentage error on comparison of CFD to experimentally
obtained results. For the windward face, all the models tested provided similar results
with a small improvement in the accuracy provided by the non-linear model. All the
models tested provided accurate peak pressure predictions at the leading edge of the
wall with the maximum errors at the far end of the wall where the CFD calculated
pressure coefficients reduced too quickly. The results for the leeward face of the wall
show accurate predictions for the maximum peak pressure, particularly so for the
RNG and non-linear models, with CFD predicted errors in the remaining profile of
approximately 20 percent. It is apparent that the flow only weakly separates on the
leeward face thus returning high negative pressures at the leading edge, as
demonstrated for the 6m cube results when using the standard k- model. Referring to
Figure 103 it is clear that this test case is less complicated than the former case as to a
large extent the flow remained attached to the wall. This is highlighted by the fact that
the results returned by the standard k- model were, in this case, reasonably accurate.
The reduced flow impingement on the leading edge and the consequent reduction in
the turbulent kinetic energy produced allowed the standard k- model to calculate a
similar flow field to that predicted by the non-linear k- model. This includes the
small leeward face recirculation shown on Figure 103.
6.8 Turbulence model: Discussion
6.8.1 The Standard k- Model
Referring to the experimental results obtained from the SRI it is clear that the standard
k- model cannot, in the majority of cases, adequately predict the flow fields and
pressure distribution around a bluff body. Nonetheless, there were a few notable
exceptions where the flow remained attached to the bluff body and flow impingement
was not so severe and the standard model performed acceptably.
105
For the 6m cube model, the most severe errors were due to the inaccurate and
excessive prediction of turbulent kinetic energy levels, particularly at the flow
stagnation point. The results of which were to raise the maximum possible windward
face stagnation pressure to a value greater than unity due to contributions from the
three normal Reynolds stress turbulence components. In addition, the excessive values
of turbulent kinetic energy predicted excessively mix and arrest the flow causing it to
remain attached to the roof of the cube. This resulted in a poor flow field prediction
and the sharp peak and large gradient of the negative pressure behind the upstream
edge. It is also apparent that excessive levels of turbulent kinetic energy and thus eddy
viscosity are generated for the majority of the flow field resulting in a surprisingly
short wake recirculation. A similar result was found for the cube skewed at 45 to the
incident wind with the excessive levels of turbulent kinetic energy reducing the
strength of the delta wing vortices.
All of the models tested, including the standard k- model, under predicted the
leeward face pressure distribution for both the normal and skewed cube. The time
averaged CFD simulations appear to predict an excessively high recirculation velocity
in the wake. Therefore the velocity deficit produced in the wake by the full-scale cube
was not matched by CFD and lower suction pressures were returned.
With regards to the results obtained from the SSB it appears that, on average, this
model performed better than for the 6m cube test. This may be due to the fact that
fewer demands are placed on the turbulence model when predicting the flow over a
sloping rather than a flat roof. This is particularly the case when considering the SSB
with curved eaves, where the flow remains attached over the entire building, with
separation only occurring in the wake.
The standard k- model performed relatively well for the Silsoe experimental wall
structure, with the worst results returned for the case with maximum flow
impingement, with the wall normal to the incident wind. As the wall was
progressively skewed to the wind and the flow impingement and separation reduced
the model provided increasingly accurate results that approached the accuracy of both
the RNG and non-linear k- models.
However, as was emphasised earlier, the standard k- model fails mainly when
modelling flows with significant impingement, due to the incorrect prediction of
turbulent kinetic energy. The reason why the standard k- model over predicts
turbulent kinetic energy in such areas is due to the eddy viscosity models inability to
106
deal with irrotational strain, which arises principally in impingement and reattachment
zones. As eddy viscosity models have arisen from, and been calibrated by, reference
flows which are strongly sheared they have great difficulty in dealing with flow fields
which are dominated by extensive straining (Leschziner 1995). Hence eddy viscosity
models which feature the turbulence energy transport equation tend to return
excessive levels of energy and thus turbulent diffusion in the presence of strong
compressive strain.
Therefore, to conclude, the errors brought about by the isotropic eddy viscosity
assumptions and the resulting simplifications, particularly to the production term in
the turbulent kinetic energy transport equation (detailed in full in sections 4.2.3.1 and
5.2.1) are primarily responsible for this models poor performance.
6.8.2 The RNG k- Model
Referring to the results obtained from the 6m cube tests, the RNG k- model (detailed
in section 4.2.6) appeared to produce relatively good results particularly when
compared to the performance of the standard k- model. It is the only widely available
turbulence model that was able to predict flow separation and reattachment on the
roof of the cube which in turn produced an improved roof pressure distribution.
Generally this model predicted relatively accurate pressure distributions for all sides
of the normal cube.
Similar improvements were found for the skewed cube, the SSB and the experimental
wall cases with relatively accurate results returned. In fact the skewed cube case
showed a significant increase in the predicted strength of the delta wing vortices,
which is attributed to the reduced turbulent kinetic energy at the flow stagnation point.
There are several reasons for the RNG k- models improvement over the results
produced by the standard model. The RNG model is derived from renormalisation
group analysis of the Navier-Stokes equations and incorporates revised model
constants used to close the transport equations. In addition, the turbulence energy
dissipation equation, which has long been considered a source of inaccuracy in both
the k- and differential stress model, has been revised. The equation now includes a
strain dependent term to aid the model in dealing with flows that experience large
rates of deformation (see Equation 4.18). These two factors are very important in
helping the model to deal with impinging flow fields and lead to reduced eddy
107
viscosity and turbulent kinetic energy production. It is stated by Orzag (1994) that the
reduced value of C
2
compared with the equivalent standard k- model coefficient has
the beneficial consequence of decreasing both the rate of production of k and the rate
of destruction of energy dissipation, leading to smaller eddy viscosities. The net effect
of this is to increase the rate of energy dissipation that will in turn reduce the value of
turbulent kinetic energy and eddy viscosity returned. Secondly, the strain dependent
term in the modified equation has the effect of modifying the rate of dissipation
depending on whether the flow is subject to isotropic or strongly anisotropic
turbulence. In areas of high straining, i.e. high turbulence anisotropy at flow
impingement areas, the value of becomes very large and the variable C
RNG
may
become negative. This will increase the production of energy dissipation in the
dissipation equation leading to further reductions in the predicted eddy viscosities.
Conversely in areas of low straining the value of and consequently C
RNG
becomes
small, thus the equation for energy dissipation will tend towards returning results
similar to the standard model equation. Therefore the net effect of these terms is to
amend the errors apparent in the standard model in areas of high straining such as
flow impingement as a consequence of the eddy viscosity concept.
Nonetheless, it should be noted that the RNG model predicted a less accurate wake
recirculation than the standard k- model. As the standard k- model generally
calculates high levels of eddy viscosity for the flow field it in turn calculates a smaller
wake recirculation. The extra terms in the RNG model reduce the levels of eddy
viscosity returned with the resulting error in the calculation of the wake vortex. This
is compounded by the fact that the RNG model lacks the ability to predict anisotropic
turbulence and underestimated the value of vv in the wake region, which in turn
leads to an underestimation of the momentum diffusion in the lateral direction
(Murakami 1993). This is in fact an important result, as it demonstrates the effect of
adding ad-hoc terms to the turbulence transport equation whereby the accuracy of
the flow predictions is increased in some areas and reduced in others. Furthermore, it
should be noted that the additions to the RNG k- model equations are not rigorously
derived from the Navier-Stokes equations and are not bounded. Consequently it has
been reported that in certain flow fields that experience very high levels of fluid
strain, the additions can destabilise the model equations resulting in convergence
difficulties and poor accuracy (Tehrani 1999).
108
6.8.3 The MMK k- Model
With reference to the 6m cube test, the general aim of the MMK k- model was to
reduce the production of turbulent kinetic energy at the front face of the cube (see
section 5.2). In this respect it achieved its aims, perhaps too well as the model
appeared to calculate a roof flow field similar to that produced with a uniform
incident wind (see Figure 5). The turbulent kinetic energy produced at the front edge
now appeared to be too small and there is no flow reattachment on the roof
whatsoever. It is also clear that the large roof recirculation had the effect of producing
a wake recirculation that extended from the base of the cube to the top of the roof
vortex. The net effect of this, combined with the low levels of turbulence and thus
eddy viscosity, produced a wake recirculation for a cube of larger dimensions than
that found by experiment. This highlights the importance of accurately predicting the
correct flow field at all points over the cube. Nevertheless an improvement over the
standard k- model can be seen for the cube windward face and roof pressure
coefficients.
Further tests were undertaken with the MMK model in an attempt to force the correct
roof flow field. These tests involved making changes to the ratio of vorticity over
shear, reducing the ratio set by Tsuchiya et al (1996) from 1 to 0.7 and 0.5. The
general idea was to increase the turbulence production value before the MMK model
was used and thus influence the roof vortex. These tests did not result in any
improvements.
The results for the skewed cube model showed a marginal improvement in the
pressure coefficient prediction around the cube and the strength of the delta wing
vortices over the standard k- model predictions. Overall it appears that although in
some areas this model produced improved results over the standard k- model, it also
produced significantly poorer results in others areas. This again leads to the general
conclusion that the use of ad-hoc modifications is unsatisfactory in wind
engineering simulations.
6.8.4 The Quadratic Non-Linear k- Model.
The results presented previously show that for a number of cases the quadratic non-
linear k- model appeared to outperform the DSM both in terms of mean pressure
distributions and predicted vortex dimensions. The fact that the more complex DSM
109
did not necessarily perform any better than the simpler models is also stated by
Mentor and Grotjen (1999). It was also clear that improved predictions have been
obtained with the anisotropic non-linear model over the isotropic eddy viscosity
models such as the RNG and standard k- models for the majority of simulations
undertaken in this study.
Referring to the 6m cube test, it has been reported by Murakami et al (1996) that the
DSM fails to predict reattachment of the roof vortex due to an underestimation of the
uw shear stress at the front corner of the cube. Although absolute values are not
available from the SRI experiments, Figures 104 and 105 show that the non-linear
model predicted markedly higher values of the shear stress at the front corner, thus
resulting in a more accurate vortex prediction. The reattachment of the roof vortex
resulted in a marginally more accurate roof pressure distribution.
Figures 106 to 108 demonstrate the abilities of the non-linear model to predict
anisotropic stresses and correctly shows the dominant stress over the roof to be the
streamwise Reynolds stress. The results for the non-linear model should be contrasted
with those produced by the standard k- model shown in Figures 109 to 111.
Although the standard k- model calculates anisotropy in the normal stresses it is clear
that the simple isotropic eddy viscosity model is incapable of accurately predicting the
stresses. Figure 109, which highlights the streamwise normal stresses, shows a
massive over prediction at the stagnation point and front corner of the cube. This is
also the case for the lateral stress component, shown in Figure 110. Figure 111 seems
to show reasonably low levels of the vertical normal stress component.
Murakami et al (1996) states that the DSM accurately predicts the distribution and
anisotropy of the Reynolds stresses over the roof of a surface mounted cube and errors
in both the eddy viscosity model and algebraic stress model are not reproduced.
Figures 112 and 113 show that there is a very similar distribution of stresses along the
roof for both the DSM and non-linear models. Furthermore, the non-linear model
correctly predicted turbulence anisotropy and higher levels of the lateral vv
Reynolds stress in the wake region, see Figures 107 and 114. However, it should be
noted that the model incorrectly predicted the streamwise Reynolds stress to be the
smallest in magnitude in the wake region. Nonetheless the results further validate the
non-linear model and highlight its improved accuracy over the ASM and eddy
viscosity models.
110
The prediction of the roof vortex size with the isotropic eddy viscosity models is
usually related to the levels of turbulent kinetic energy produced at the front face
stagnation point. Consequently it was interesting to note that although the non-linear
model predicted a lower level of kinetic energy than the RNG model it still predicted
a shorter more accurate roof vortex. This was most likely due to the increased lateral
diffusion predicted by the model, due to its anisotropic turbulence modelling abilities,
and the increased levels of shear stress returned as discussed above. The same
arguments also apply to the wake region. Improvements over the standard k- were
also apparent for both the velocity and pressure distributions obtained from the
skewed cube, although it should be noted that no improvement is apparent over the
RNG model for this case.
The results from the Silsoe experimental wall and SSB experiments further
demonstrated the non-linear model capabilities, although it should be mentioned that
the latter results were generally poor for all the models tested for a number of the
measurement locations.
Non-linear k- models have been developed in an attempt to incorporate the effects of
anisotropic turbulence and thus to try to obtain a width of applicability approaching
that of the differential stress model. The advantage of using non-linear expansions of
the Boussinesq hypotheses to develop improved turbulence models is that it should
produce universal improvements. This is in contrast to ad-hoc improvements, such as
the MMK model that may improve predictions only for a particular application of the
model. To a much lesser extent, the wake results obtained from the RNG model have
further highlighted this. The overproduction of turbulent kinetic energy, as explained
in the k- model results, is not an issue with this model, which uses the full turbulent
kinetic energy production term. Furthermore, non-linear models require much less
computational effort, due to their improved stability characteristics, as demonstrated
in this work.
6.8.5 The Differential Stress Model.
The DSM, which solves an extra transport equation for each of the Reynolds stresses,
is a fully anisotropic turbulence model. Each of the transport equations for the
Reynolds stresses incorporates terms to deal with curvature related strain, adverse and
favourable pressure gradients and other flow effects. Consequently when taking
account of the greater modelling abilities of this more complex model it was rather
111
surprising to find that it does not predict a very accurate roof flow field for the 6m
cube. The initial conclusion is that this error must be attributed to the lack of wall
reflection terms in the pressure-strain model. However, Murakami et al (1993) has
previously undertaken testing of a total of 6 differential stress models which
incorporated different pressure-strain models for the flow over a surface mounted
cube in an atmospheric boundary layer. Although a scale model was used, this
simulation used the same general test case and flow field as the tests undertaken in
this work. Therefore the physics and general effect of the extra terms will be identical.
The results showed that there was no improvement in the prediction of the roof
vortex. It is apparent from this that although these terms are important for certain flow
fields, they are perhaps of relatively small importance for wind engineering flow
fields that are dominated by the convection and diffusion of large scale eddies and
coherent structures in the flow, as discussed in section 5.3.
Referring to the SSB simulations the DSM appeared to perform in a similar manner to
the normal cube tests with similar accuracy to the RNG k- model. Furthermore, the
significant instabilities and convergence difficulties experienced when using the DSM
with the 6m cube model, were further compounded with the SSB geometry. The
general instability of this model resulted in the failure to successfully apply the DSM
to any transverse SSB cases. Therefore these tests highlighted the inability of the
DSM to provide converged results for a marginally more complex geometry than the
previous test case and thus its failure to meet one of the most important criteria for a
suitable wind engineering turbulence model. It could well be argued that the RNG
model calculated a comparatively accurate pressure and flow field distribution at a far
lower computational cost. Nevertheless the DSM provided encouraging
improvements for the skewed cube case, which has traditionally been the cause of
major errors with eddy viscosity models.
6.8.6 Results summary
The overall summary of the results of the testing, based on the criteria stated in the
introduction (section 6.1) are given in Table 8.
112
Turbulence model Accuracy Stability Comp.
overheads
Standard k- Low Medium-high Low (1.2)
RNG k- Medium-high High Low (1)
MMK k- Medium Medium Low (1.2)
Quad non-linear k- (Speziale 1987) ------- -------- ---------
Quad non-linear k- (Craft et al 1996) High High Low (1)
Cubic non-linear k- (Craft et al 1996) * High Very low Low (1)
DSM Medium-high Low High (3)
Table 8: Results summary (* limited data available)
Note the assessments given are relative to the other turbulence models tested. The
numbers given in the computational overheads column refer to the comparative run
times (in days) for the different turbulence models, using the same model test case,
when ran on a Digital Dec-Alpha Unix workstation.
Overall the accuracy of all the models tested in these wind engineering flow fields
was disappointing. This was particularly the case for some of the measurement points
on the SSB, which returned CFD prediction errors as high as 100 percent.
113
7. Large Eddy Simulation (LES).
7.1 Introduction
The subject of large eddy simulation is at present too important a technique not to
investigate although, unfortunately, it is too large a subject to investigate fully as part
of this thesis. The application of LES to computational wind engineering is worthy of
several theses and therefore the aim of this chapter is to undertake a limited number of
LES tests and to inform the interested reader of future work that should be pursued in
this field.
The aim of this analysis was mainly as a preliminary study to both assess the
difficulties in applying this technique to wind engineering and the minimum hardware
requirements necessary. Interesting transient details omitted from the steady state
simulations, undertaken in chapter 6, will be presented. The overall conclusions
gained have been based on the validity of this model as a wind engineering technique
in terms of its accuracy and applicability. The results of the LES work have also been
compared to a number of steady state models simulations, including the low Reynolds
number k- and k- and the high Reynolds number quadratic non-linear k- models.
All of these simulations have used the same model geometries with periodic inlet and
outlet conditions to provide basic comparative data.
The following sections that detail the theory behind large eddy simulations are based
on the work of Jacobsen (1997). The methods used to implement and test the LES
technique, using CFX4.2, are based on the work of Jacobsen (1997) and Ciofalo and
Collins (1989).
7.2 Theory: Filtering the Navier-Stokes Equations
With large eddy simulations it is essential to define those quantities that are to be
computed precisely. A velocity field is required that contains only the large scale
components of the total field. This is best achieved by filtering the large or resolved
scale field (Leonard 1974). As such, in large eddy simulation, any physical quantity
f is decomposed into two parts:
" f f f + (7.1)
114
Here f is the resolvable scale component and " f is the sub-grid-scale (SGS)
component. Assuming that " f is the filtered velocity,
i
u , then it is defined with a
filter function ) ' , ( x x G as follows:

' ) ' ( ) ' , ( ) ( dx x u x x G x u


i i
(7.2)
The filter kernel ) ' , ( x x G is a localised function and may include a Gaussian, box or
cutoff filter. Every filter has a length scale associated with it that dictates which
eddies are large and are directly resolved and which are small and therefore modelled.
Imposing the filter on the continuity and Navier-Stokes equations produces the
following filtered equations:
0

x
u
i
(7.3)
j j
i
i j
j i
i
x x
u
x
P
x
u u
t
u

2
. .
1

(7.4)
where the over-bar means a space filtered quantity.
In order to define the large scale velocity field that is separated out from the sub grid
scale components a Reynolds averaging type procedure is carried out:
i i i
u u u ' + (7.5)
As was described in section 3.2, the main difficulties are due to the non-linear
convection terms in the Navier-Stokes equations.
Putting Equation 7.5 into the non-linear terms and expanding gives:
( )( )
(4) (3) (2) 1) (
+ + +
+ +
j i j i j i j i
i i j j i j
u u u u u u u u
u u u u u u


(7.6)
In general it must be noted that
l k
u u ' ' 0 . In contrast to the time averaged approach
all the four terms must now be considered. Terms 2 to 4 contain small scale, non-
115
resolvable components
i
u' , which must be modelled. Term 1 is defined entirely in
terms of the grid-resolved quantities by introducing the SGS stresses (Jacobsen 1997):
j i j i ij
u u u u (7.7)
inserting this into Equation 7.6 means that the convective term in the filtered Navier-
Stokes equation, Equation 7.4, will only depend on the resolved quantities (Jacobsen
1997).
7.3 Sub-Grid Scale Stresses
The sub-grid scale stresses can be expressed as the sum of three terms as follows:
ij ij ij ij
R C L + + (7.8)
where
j i j i ij
u u u u L (7.9)
j i j i ij
u u u u C
' '
+ (7.10)
' '
j i ij
u u R (7.11)
ij
L are referred to as Leonard stresses,
ij
C are the cross terms and
ij
R are the SGS
Reynolds stresses.
These stresses are modelled in the following manner:
Leonard (1974) has shown that the Leonard stress term removes significant
energy from the resolved scales. They can be computed explicitly but it has
been shown that they are of the same order as the truncation error when a
finite-difference scheme of order higher than two has been applied and thus
they are implicitly represented (Wilcox 1994). The remaining stresses have
been modelled in different ways. Their properties are assumed to be identical
with those arising in the Reynolds time-averaging approach. For the sub-grid
scale turbulence the eddy viscosity model assumes that the sub-grid scale
116
stress is proportional to the modulus of the strain rate tensor of the filtered
large-scale flow (Jacobsen 1997 p 48).
Therefore the SGS eddy viscosity is modelled using the same analogy as that
proposed by Boussinesq and detailed in chapter 4:
i
j
j
i
sgs ij sgs ij
x
u
x
u
S

2 (7.12)
In a similar manner to the eddy viscosity models described in chapter 4 an extra term
is added to this equation to model the effect of the SGS turbulence kinetic energy and
normal stresses
ij ij sgs ij
k S
3
2
2 (7.13)
This sub-grid scale stress model effectively models the sub-grid viscosity. It plays a
role in LES similar to the role played by the Reynolds stresses in the RANS models
but the physics that it models are different. The SGS energy is a much smaller part of
the total flow than the RANS turbulent energy and thus model accuracy may be less
crucial in LES than RANS computations.
7.3.1 The Smagorinsky Model
This model, proposed by Smagorinsky in 1963, is based on an equilibrium hypothesis:
that the production of SGS energy is approximately in equilibrium with the viscous
dissipation of the SGS energy. This gives:

ij ij
S (7.14)
It can be shown that the sub-grid scale eddy viscosity
sgs
then relates to the strain
rate tensor of the resolved velocity field
ij
S (Piomelli 1994).
[ ]
2
1
2 2
2 ) ( ) (
ij ij s s sgs
S S C S C (7.15)
The model constant
s
C can be varied between 0.1 and 0.2 and affects the sub-grid
viscosity and therefore the rate of energy drainage from large to dissipative scales.
117
Since an eddy dissipation formulation for the sub-grid scale stresses is used it is
thereby assumed that the sub-grid turbulence is isotropic. As the smaller scales of
turbulence are more likely to be isotropic due to the damping effects of viscosity, this
can be considered to be a valid assumption. Therefore the filter width can be found for
a Cartesian grid to be (Jacobsen 1997):
3
1
) (
z y x
(7.16)
These simulations used an anisotropic grid distribution to reduce the overall number
of finite volume nodes required in the simulations. There have been a number of
investigators who have attempted to account for the effect of an anisotropic grid on
the filter width and Smagorinsky model, but it has been reported that these
modifications only showed small deviations from the standard isotropic grid method
(Scotti and Meneaveau 1993).
The standard Smagorinsky model is so simple and well designed that it has been
applied to many flow fields and has attained great success. However the standard
model has several shortcomings:
1. the model is overly dissipative
2. the Smagorinsky constant C
s
must be optimised for each flow field
3. no natural account is taken of the effects of walls
4. no account is made of the effects of negative SGS viscosities or the apparent
transfer of energy from the small to large scale eddies, known as backscatter (see
section 5.7).
The second shortcoming represents the most serious problem in wind engineering
flow fields. The rate of transfer from the large eddies to the small dissipative eddies
varies according to both the spatial and temporal point in the flow field as a
consequence of the many different flow types encountered. Consequently dynamic
SGS models have been devised that calculate values of the model constant dependant
on the spatial and temporal conditions in the flow field. For full details see Germano
et al (1991).
118
7.4 Numerical Methods The Time Dependant Terms
The CFX user has a choice of either a fully implicit backward differencing scheme or
a time centered Crank-Nicolson treatment that is second order accurate in time.
A particular advantage of using the Crank-Nicolson scheme is apparent when the
Courant-Friedrichs-Lewy or CFL criterion is adhered to obtain a stable solution. The
CFL criterion states that:
0 . 1
.
<

x
t u
N
CFL
(7.17)
If the implicit Crank-Nicolson method is used the CFL can then be exceeded by
factors as large as 5 (Wilcox 1994) thus allowing for larger time-steps for the same
model simulation. Trials undertaken by the author have generally agreed with the
findings of Choi and Moin (1994) in that too large a time step will effectively damp
the turbulent fluctuations leading to a laminar solution. Further details of the schemes
highlighted can be found in Patankar (1980) and the CFX version 4.2 user manual
(CFX-International 1998).
7.5 Model Validation and Analysis Details
The standard Smagorinsky LES model has been incorporated into CFX through the
appropriate user FORTRAN routines and has been validated through previous tests
undertaken by Ciofalo et al (1989) and Jacobsen (1997) as part of their respective
Ph.D. theses. This model has been chosen due to its simplicity and ease of
implementation into CFX4.2. The use of an anisotropic grid and the method of
generating the inlet conditions, discussed below, have been validated through
comparisons with the work of Scotti and Meneaveau (1993) and Thomas and
Williams (1992), respectively. All the simulations undertaken were run to single
precision numerical accuracy to speed up the analysis times and reduce the random
access memory requirements.
The following section details the relevant calculations required to undertake the large
eddy simulations and any assumptions made.
119
7.5.1 Domain Size
For the tests undertaken in this project a number of different geometries were used as
follows:
1. 1:30 scale, cube model, normal orientation, fine mesh with approximately 1
million nodes and Van Driest wall damping (see section 7.5.5)
2. 1:30 scale, cube model, normal orientation, a mesh with approximately 180,000
nodes, no Van Driest wall damping
3. 1:30 scale, SSB model, sharp eaves, normal orientation, fine mesh with
approximately 1 million nodes and Van Driest wall damping.
Due to time constraints it was not possible to undertake further grid refinements.
Strictly speaking the use of the standard Smagorinsky SGS model requires near wall
damping to reduce the increased levels of eddy viscosity returned in the near wall
region (see section 7.5.5). The necessity for wall damping depends to a large extent
on the type of flow that is being simulated. Wind engineering flows are mainly
concerned with the effects of the large scale structures in the flow so the errors in the
Smagorinsky model in the near wall region are perhaps not a major concern.
Nevertheless it was felt necessary to use a geometry that correctly modelled the near
wall region by placing at least one grid point in the laminar sub-layer. Therefore a
number of different tests were set up, as detailed above, to test this hypothesis.
Using the 6m model and the geometry dimensions shown in Figure 115 it was not
possible to mesh the domain for a 6m cube, which was approximately 84m by 42m by
18m, to allow for near wall damping functions without the use of an unacceptably
high number of grid points or a very large geometric progression of the grid. The 1/30
scaling of the model reduced the domain size to 2.8m by 1.4m by 0.6m, and as the
sub-layer thickness is the same for a given flow Reynolds number and wall roughness,
it was therefore the same for both these models. It was now possible to economically
mesh this domain for test numbers 1 and 3.
120
7.5.2 Inlet conditions
Further problems were apparent due to the fact that no transient boundary layer data
was available from the SRI for use as inlet conditions in these simulations. This was
further compounded by the fact that the rough wall functions used in the turbulent
steady state simulations were not available when the laminar flow settings were used
in CFX4.2. Due to time constraints on this work it was considered appropriate to
generate transient inlet boundary conditions using periodic inlet and outlet boundaries
to recycle the flow and use a non-slip, smooth wall condition for the domain floor.
The initial conditions for the mean velocity profiles at each cell were set to obey the
power law expressed as:
25 . 0
Z u (7.18)
The initial velocity fluctuations for the u and v components were generated by means
of random numbers. The turbulence intensities were set to 15 percent and the initial
velocity fluctuation for w was set to satisfy the continuity equation (Murakami et al
1987). After many cycles the initial velocity fluctuations were effectively overwritten
as the turbulence in the flow evolves. These initial conditions effectively speed up the
process of generating turbulence rather than starting the simulation from a laminar
flow.
Validation of the method of generating the inlet conditions is available on reference to
Thomas and Williams (1992) and Castro and Robins (1977). It is suggested that
turbulence levels induced by the cube above the background turbulence extend to
approximately 6 times the stream-wise cube dimension (Castro and Robins 1977).
Therefore provided the computational domain is longer than say 8 times the cube
dimension it is possible to generate the inflow by recycling the outflow (Thomas and
Williams 1992). The inflow parameters are therefore not set a priori but are
determined by the simulation and the turbulent boundary layer is generated as a
consequence of the non-slip floor wall condition set, as opposed to the turbulence
generated by the cube. A pressure drop across the domain was specified in order to
maintain the flow through the computational wind tunnel. General guidance on how
to estimate the resulting wind velocities is given by Ciofalo (1989). Using this work
and, to an extent, some trial and error testing the velocity and Reynolds number for
these simulations measured at eaves height was approximately equal to 1.7 m/s and
28,000, respectively.
121
7.5.3 Determination of Time-Step
The appropriate time step was determined using the simple CFL criterion discussed
earlier (section 7.4) with the worst case combination of streamwise velocity and
smallest cell size for the various test cases. Consequently a number of different time
steps were chosen, depending on the model geometry used, ranging from
approximately 0.01 to 0.0001 seconds
7.5.4 Large Eddy Turnover Time
The simulations have to be run for a period that ensures that the turbulent structures
have fully developed before any averaging can be performed (Jacobsen 1997). This
period is a multiple of the large eddy turnover time (LETOT) which refers to the time
it will take for the large eddies in the simulation to do one revolution.
For this case the largest eddies in the flow are of approximately the size of the
building, consequently the time for one LETOT is as follows for test case number 1:
LETOT = Building dimension, (H) / velocity at roof height (7.19)
Therefore one LETOT = 0.2/1.7 = 0.12 seconds.
It was necessary to run the simulations for a time that ensured that sustained and
statistically stationary turbulence levels were generated beyond the initial conditions
set. This required approximately 20 to 30 LETOTS prior to the averaging of the
results (2.35 to 3.52 seconds).
7.5.5 Wall Conditions
As was briefly highlighted in section 7.5.1, obtaining satisfactory wall conditions
when undertaking large eddy simulations is a major problem when using the
Smagorinsky SGS model. It has been discovered that the conventional Smagorinsky
model returns increased values of sub-grid scale viscosity in the near wall region. This
is in contrast to the fact that viscous effects give way to laminar flow below
+
y
values of approximately 11.22. Consequently an ad-hoc method is required when
using this sub-grid model to force it into agreement with experimental results. A
122
recipe that has been found to be successful is Van Driest damping that has long been
used to reduce the near wall viscosity in RANS models. The Van Driest damping
factor, which is normally used, is as follows (Driest 1956):

,
_


+
+
A
y
d exp 1 (7.20)
where d is the damping factor used to modify the SGS eddy viscosity returned and A
+
is a constant equal to 25.
The wall distance in wall units is found from:

u
y y .
+
(7.21)
where

wall
u is the wall friction velocity. (7.22)
In addition,
It should be noted that the expression for the wall friction velocity is strictly
only valid for a channel flow without recirculation zones including separation
and re-attachment points. This is due to the fact that
wall
and the wall friction
velocities vanish at reattachment points. Nevertheless this method is
implemented in the absence of a more advanced solution to the problem
(Jacobsen 1997, p55).
The wall shear stress is found from the dynamic viscosity multiplied by the gradient
of the main streamwise velocity parallel to the wall:
y
u
wall

(7.23)
A similar procedure was used as with the steady state RANS models (described in
section 6.3.8) to test that a satisfactory number of grid nodes were in the laminar sub-
layer. Although this modification is difficult to justify in the context of LES it does
produce the desired results and has been successfully applied to many engineering
flow problems.
123
7.5.6 Calculation of Velocity Statistics
In order to directly compare the large eddy simulation data with the mean data
provided by the steady state, low Reynolds number turbulence models it was
necessary to time average the velocity statistics. This involved calling a user
FORTRAN routine at each time step that calculated the mean velocities and root
mean square fluctuating velocities as follows:
start
t
tstart
n
t t
u
u

(7.24)
start
t
tstart
n
t t
u u
u

2
2
) (
' (7.25)
The time averaging process started after a designated number of time steps, t
start
, to
allow the solution to develop and generally the longer the time averaging process the
more accurate the calculated mean and RMS values became (Jacobsen 1997). Using
the information found from Equations 7.24 and 7.25 it was also possible to calculate
transient and time averaged Reynolds stress data.
7.6 Results
7.6.1 Introduction
It is intended that the analysis of the results will focus primarily on the flow fields
produced by the different models used and a discussion will be given on any
difficulties found during these simulations. The discussion initially focuses on the
tests undertaken to assess whether the LES simulations are performing correctly. This
is followed by analysis and comparison of the steady state results from the low
Reynolds number and the quadratic non-linear k- models and the LES data.
124
7.6.2 Steady State Simulation and LES Results
Various checks were made using the LES models to test whether a suitable level of
turbulence had been generated and that the LES algorithms incorporated into CFX
were performing correctly.
It is stated that the turbulent flow over a surface mounted bluff body in an
atmospheric boundary layer is only quasi-periodic or even completely non-periodic
(Shah and Ferziger 1997). This can be seen upon reference to experiments undertaken
with a square cylinder, refer to Frank and Mauch (1993) for further details.
Following the method used by Shah and Ferziger (1997) a probe was positioned at
the side wall of the cube model to return fluctuating velocity data, for a single node,
for each time step of the simulation. Figure 116 shows the time series of the
fluctuating velocity returned. The change in direction of the flow, highlighted by the
time series, is graphically demonstrated by Figures 117 to 120 that show snapshots of
the flow generated by the coarse model at 2-second intervals. The flow appears to
change direction as the stagnation point at the front face of the cube switches from
either the left or right of the centre of the front face, which results in the flow either
separating at the side or remaining attached.
The Fourier transform of the time series signal (see Figure 121) showed no peak in
the spectrum. This matches the results found by Shah and Ferziger (1997) who state
that this is typical of turbulent flows where repeatable structures are found but they
are not identical in size or strength and do not occur periodically. As LES is unable to
adequately resolve the small scale eddies responsible for dissipation the decay in the
energy of the flow with increasing frequency (see Figure 121) represents the effects of
the SGS model and near wall damping terms. As such any assessment of this region
of the spectrum is particularly difficult.
Referring to Figures 122 to 124, that show snap-shots in time of the flow for various
planes for both the SSB and cube models, it is apparent that the flow appears very
different from the steady state simulations tested and is in fact highly turbulent and
transient. Unfortunately these fine grid simulations only had limited success due to
the need to finely mesh the near wall region of the flow and the use periodic
boundaries for the inlet and outlet. The result of these constraints meant that it was
necessary to geometrically progress the grid to the outlet to allow for a model that had
at least a computationally workable finite volume mesh. This technique led to a
125
situation where the turbulence generated in the flow was effectively damped by a
finite volume mesh that became progressively coarser as the outlet was approached.
Both the LES cube and SSB model simulations returned time averaged flow fields
that appeared to show a reduced level of turbulence in the flow on comparison to the
steady state models, i.e. no flow reattachment on the roof of the cube. Unfortunately
continued run time for these simulations showed no further increase in turbulence
levels. This is demonstrated in Figure 124, which shows the transient flow field for
the entire SSB domain.
As a consequence of this and in the knowledge that, reportedly, wind engineering
simulations are less sensitive to the near wall treatment (Hanjalic et al 1998), a
number of coarse model simulations were undertaken without Van Driest wall
damping that allowed for a more uniform grid. Figures 125 and 126 show the level of
turbulence generated in the flow, for this coarser model, which, on comparison to
Figure 124 for the SSB, appears to be better sustained when convected downstream.
Figures 127 and 128 show the time averaged velocity and pressure contours around
the cube.
Using the information gained from the various checks undertaken it was possible to
obtain suitable time averaged LES data to compare to the steady state turbulence
model results.
Referring to Figure 127 which highlights the roof and wake recirculation zones for the
LES simulations and Figures 129, 130 and 131, which show the recirculation zone
data for the time averaged turbulence model runs, the following results were found:
Turbul ence model . Roof reattachment. Wake reattachment.
Non-l i near k-epsi l on. 0. 67H 2.3H
Low Re. No. model s. No separat i on 2.2H
LES approx. 0. 6H 1.8H
Table 9: Reattachment lengths, cube normal to the incident wind.
(H = cube dimension)
Interestingly with this simulation it appears that a similar flow field was being
produced, by the non-linear k- model, to that modelled in chapter 6 for the 6m cube.
The flow field results from the low Reynolds model simulations, which were identical
for both the k- and k- models, show no improvement on the results found for the
standard k- model tested in chapter 6.
126
Although the large eddy simulation perhaps required a longer averaging time, the
results were sufficient to determine some important flow data. The time averaged
coarse LES model returned a similar roof flow field to that predicted by the non-linear
k- model and a wake reattachment length of approximately 1.8 times the cube length.
Although only tentative conclusions can be drawn from the data presented in this
section of the thesis, it should be stated that inaccuracies in the calculation of the
wake flow field, which resulted in over prediction of the reattachment point, were one
of the consistent and unresolved problems of the simulations undertaken in chapter 6.
The early results from the large eddy simulations were therefore generally
encouraging.
7.6.3 Discussion and Recommendations for Future Work
The scope of the LES work undertaken in this section is just the starting point of a full
investigation into LES and wind engineering. The work undertaken has given the
author an insight into actual LES work beyond simply researching published material
on this subject and has tentatively highlighted possible improvements in predictive
accuracy.
This initial approach has conclusively shown the requirement for future work in this
field. Firstly in order to compare the SRI cube results to any LES data it is necessary
to obtain transient boundary layer data from the field experiments and convert this
into data to be used at the inlet for the CFD simulations. Secondly a non-trivial
amount of work needs to be undertaken to supply suitable wall functions that will
sustain the inlet boundary conditions and match the roughness of the fetch at the SRI.
The author is of the opinion that these problems alone could well be the basis of a
further thesis in their own right. Thirdly SGS models need to be incorporated that rely
less heavily on the ad-hoc damping of the near wall eddy viscosities returned, perhaps
utilising the dynamic SGS model of Germano (1991). Fourthly, and perhaps most
importantly, is the issue of the necessary computational power and data storage
facilities. Even the limited simulations undertaken as part of this project required the
use of some expensive computer hardware (multi-processor Sun Unix workstations)
that turned out to be barely adequate for this project with run times of up to 6 months
for a number of the LES models. More detailed studies may require more substantial
hardware or parallel processing facilities. This alone implies that LES will not be
routinely used in the near future.
127
Finally this section has clarified that, at least for the next 5 years, most CFD
simulations will be undertaken using the techniques utilised in chapter 6. Not only
does LES require major computer hardware it also requires significant operator skills.
LES appears to be a much more difficult, less black box, approach to modelling
fluid flow with a much greater scope for error. Significant operator skills and a good
knowledge of fluid mechanics are required.
128
8. Conclusions and Recommendations for Future Work.
8.1 Introduction
This project has seen the first exhaustive test of all the available non-linear k- models
and the RNG k- model in wind engineering simulations with validation data supplied
from full-scale rather than model scale experiments.
In this chapter a number of conclusions are drawn from the work undertaken and to
aid clarity they will be presented in list form.
8.2 Conclusions
Primary
Development work on a commercial CFD code has seen the successful
implementation and testing of state of the art methods of modelling turbulence and
fluid flow. Extensive tests undertaken on the non-linear k- turbulence models
have shown that they produce, on average, the most accurate pressure and flow
field predictions of all the models tested in wind engineering applications. In
addition the Craft et al (1996) quadratic model is numerically stable and has
solution times equal to that of the standard k- model. Overall, in terms of
predictive accuracy, the RANS turbulence models are ranked as follows:
1. The cubic and quadratic non-linear k- models
2. The RNG k- model / DSM
3. The MMK k- model / standard k- model.
These investigations suggest that a suitable turbulence model for wind engineering
should be able to model the anisotropy of turbulent flow, such as the Reynolds
stress model, whilst maintaining the ease of use and computational stability of the
two equation models. This is necessary, firstly to account for the fact that wind
engineering flows are inherently anisotropic and secondly to aid correct
calculation of the turbulent kinetic energy production term, which has been shown
to be the main reason why the isotropic eddy viscosity models fail. The
requirement for ease of use and computational stability has been highlighted by
129
the numerical stiffness problems associated with the DSM. It appears that, at least
for the next five years, RANS techniques, such as the non-linear models, offer the
most promising computational models in wind engineering.
It should be stated that none of the RANS models provided accurate results for the
whole range of test cases, with errors sometimes exceeding 100 percent. This
suggests the eventual need to reduce the importance of the turbulence model and
the necessity to resolve more of the flow field directly, i.e. large eddy simulation.
Unfortunately, as was highlighted in chapter 7, large eddy simulation currently
requires significant computer facilities. In the authors opinion this will hold back
the widespread use of such a technique for between five to ten years after which
time this method will most likely be the dominant CFD technique used.
Secondary
A number of points regarding the use of the more popular turbulence models and
various methods of refining the models have been discovered as follows:
1. The inadequacies of the standard k- model in wind engineering flows has
been extensively shown and discussed. The problems are rooted in the
isotropic eddy viscosity concept and the linear Boussinesq hypothesis
which not only result in an inadequate description of turbulence but also
major errors in the turbulence model transport equations.
2. The methods referred to in this thesis as making ad-hoc modifications to
turbulence models, such as the MMK k- model, are generally
unsatisfactory. Although this technique was found to improve the results
in some areas of the flow it had the opposite effect in other areas. Care
should be exercised when using such models to ensure that results are
checked over a wide range of flow fields. This method is generally
unsound and modifications should be rigorously derived from either the
Navier-Stokes equations or Boussinesq hypothesis.
3. The strongly coupled Reynolds stress partial differential equations result in
finite difference equations that are highly numerically unstable. Particular
problems are apparent when the differential stress model is used in
conjunction with higher order differencing schemes. These problems were
130
so severe as to negate its use for a number of simulations in this project.
Use of the differential stress model in complex flow fields requires skill
and expertise to coax a fully converged solution. In addition this model
requires greater overheads in terms of time and computational power.
Nonetheless, it is the authors opinion that the DSM is, in theory, a rational
and sound model which may, following some major improvements,
become a useful wind engineering turbulence model in the future. Further
development work on this model should concentrate on improving the
stability characteristics and accurately pinpointing the errors in this model.
Further supplementary points regarding the turbulence model developments are as
follows:
1. At present the Speziale (1987) quadratic and Craft et al (1996) cubic non-
linear k- models appear to require further development work to improve
their numerical stability in wind engineering applications. Simple tests
have highlighted the greater accuracy and potential of the cubic non-linear
model.
2. Particular errors were apparent in modelling the flow over more complex
structures such as the SSB and the 6m cube skewed to the incident wind.
The problems associated with CFD and turbulence modelling in correctly
predicting the delta wing vortices and lobes of high negative pressure still
persist, although it is possible that difficulties in flow resolution at these
areas may be partially to blame.
3. The problems associated with the wall reflection terms in the pressure-
strain model appear to have no solution for the foreseeable future. At
present the use of such terms makes the DSM unusable in a generic sense
and as such of no use to the wind engineering community as a whole. This
is a non-trivial issue that is at present holding back the development and
use of a complete DSM.
Further points regarding the CFD testing in general are as follows:
The experimental data supplied by the SRI for full-scale structures has enabled
both novel and high quality analysis of the accuracy of state of the art CFD
turbulence models when applied to wind engineering studies.
131
The first full application of the differential stress model to simulating the flows
around the experimental structures at the SRI has been completed. In addition, full
mean Reynolds stress inlet profiles, for a range of velocities, have been generated
that represent the atmospheric boundary layer at the SRI site.
This work has clearly shown the need to undertake a significant number of tests to
determine the accuracy of a CFD simulation. In particular the importance of
undertaking the following has been demonstrated:
1. successive grid refinements to prove a grid independent solution.
2. the importance of accurately specifying the inlet conditions for any
CFD wind engineering problem
3. that all CFD simulations should incorporate higher order differencing
schemes such as CCCT or QUICK to minimise the effects of false
diffusion
4. and when solving flows that involve complex flow fields and
geometries the use of algebraic multi-grid methods may be beneficial.
8.3 Recommendations for Future Work.
The full potential of the non-linear k- models should be realised by development
work on the cubic variant to improve its stability in wind engineering flows. The
author believes this will be the most rewarding path to take rather than
concentrating on finding the highly elusive universal wall reflection term, for the
differential stress model, that is not dependent on the wall normal distance. The
DSM presently has a number of other problems, as described, which should take
precedence.
Full SRI transient inlet conditions and suitable wall functions need to be
developed to enable a complete assessment of the ability of LES to model fluid
flow for the full-scale structures tested at the SRI. Analysis and back-
engineering of presently available frequency spectra obtained from the full-scale
132
cube, SSB and experimental wall studies may provide a possible solution to this
problem.
More detailed experimental work is required to further validate the inlet
conditions for the differential stress turbulence model.
133
Nomenclature
+
A = a constant equal to 25.
ABL = Atmospheric Boundary Layer
C
1,
C
2,
C
1RNG,
C

= constants in the turbulence transport equations
C
ij
= convection term of ' '
j i
u u
s
C = model constant
CFD = Computational Fluid Dynamics
CWE = Computational Wind Engineering
DSM = Differential Stress Model
d = Van Driest damping factor
D
ij
= diffusion term of ' '
j i
u u
ij
e = the rate of deformation tensor
F = body force per unit volume.
F
ij
= stress production of ' '
j i
u u by action of rotational or body forces
k = turbulent kinetic energy, k = ( )
, ,
2
1
i i
u u
l = length scale
LES = Large Eddy Simulation.
LETOT = Large Eddy Turn Over Time
N
CFL
= the Courant-Friedrichs-Lewy or CFL number
P = mean pressure
P
k
=

production of turbulent kinetic energy
ij
P = the production term in the algebraic stress model
RANS = Reynolds Averaged Navier-Stokes
s = shear
SGS = Sub-Grid Scale
S
ij
=

strain rate tensor,

,
_

i
j
j
i
ij
x
u
x
u
S
S
(n)
= spectrum function.
SSB = Silsoe Structures Building
t = present time step
t
start
= time step at start of averaging process
velocity. of component g fluctuatin random =
ity. mean veloc =
u
u

u
*
= friction velocity
+
u = velocity scale (non-dimensional)
i
u = the filtered resolvable scale velocity field, and
134
i
u' = the sub-grid scale (SGS) components.
' '
j i
u u = Reynolds stresses
U
ref
= reference velocity
U z ( ) = the mean velocity at height z
x
i
= the co-ordinate direction (subscript indicates the coordinate direction,
x, y and z for i, j and k)
+
y = the distance from the wall in non-dimensional wall units
y = the distance from a point in the fluid to the nearest wall
z = vertical height
z
0
= ground roughness length
Z
ref
= reference height
= gradient height
ij

=

Kronecker delta
= the filter width.

t
= time step size or duration.

x
= grid spacing.
= dissipation rate of k

ij
= dissipation term of ' '
j i
u u

ij
= pressure-strain correlation term
= Von Karmans constant

l
= laminar viscosity
= dynamic viscosity.

t
=

eddy viscosity
= the laminar kinematic viscosity.
sgs
= the sub-grid viscosity
= fluid density

k ,

=
turbulent Prandtl number, for k and , for the k- turbulence model
= shear stress
ij
= the turbulent stresses tensor
O
= surface shear stress
= vorticity (equal to twice the instantaneous rate of spin of a fluid
element, determined by lateral velocity gradients.)
135
References.
Abbott, M. B. and Basco, D. R. (1989). Computational Fluid Dynamics - An
introduction for Engineers., Longman Scientific, Singapore.
Abou-Arab, T. W. and Serag-Eldin, M. A. (1992). Turbulence modelling and
simulation of atmospheric boundary layers., Transactions of the ASME, Vol. 114,
No.40, March 1992.
Anderson, J. D. (1995). Computational Fluid Dynamics - The basics with
applications., Mac Graw - Hill Inc. , New York, USA.
Baetke, F., Werner, H. and Wengle, H. (1990). Numerical simulation of turbulent
flow over surface mounted obstacles with sharp edges and corners., Journal of Wind
Engineering and Industrial Aerodynamics, Vol. 35, pp. 129-147, Elsevier,
Amsterdam, Netherlands.
Baker, C. J. (1996). Wind Engineering course notes, Master of Research Degree, The
University of Nottingham , Nottingham, UK.
Baker, C. J. (1997). Private Communication, The University Of Nottingham,
Nottingham, UK.
Basora, B. and Younis, B. A. (1992). Progress in the prediction of turbulent wind
loading on buildings., Journal of Wind Engineering and Industrial Aerodynamics,
Vol. 41-44, pp. 2863-2874, Elsevier, Amsterdam, Netherlands.
Bernard, P. S. and Speziale, C. G. (1992). Bounded energy states in homogeneous
turbulent shear flow - An alternative view., Journal of Fluid Engineering, Vol.114,
pp. 29, March 1992.
Bickerton, J. E., Graysmith, J. L., Howell, J. P. and Haynes, T., (1996) The CFD
investigation of flow separation over a simple vehicle model., ISATA, Florence,
Italy.
Boussinesq, T. V. (1877). Mem. Pre. Acad. Sci., 3
rd
Edn., Paris XXIII, p. 46.
Bradshaw, P. (1973). Effects of streamline curvature on turbulent flow.,
AGARDograph no. 169.
136
Bradshaw, P. (1992) Turbulence: The chief outstanding difficulty of our subject., 5th
Symposium on Numerical and Physical Aspects of Aerodynamic Flows, January
1992, California State Univ., USA.
BS6399 part 2. (1997), Code of Practice for Wind Loads., HMSO, London, England
Castro, I. P. and Robins, A. G. (1977). The flow around a surface mounted cube in a
uniform and turbulent shear flow., Journal of Fluid Mechanics, Vol. 79, part 2, pp.
307-335.
CFX - International (1998). CFX Version 4.2: User Manual., Harwell Laboratory,
Oxfordshire, OX11 0RA, UK: CFX - International.
Choi, H. and Moin, P. (1994). Effects of the computational time step on numerical
solutions of turbulent flow., Journal of Computational Physics, Vol. 113, part 1, pp.
1-4.
Ciofalo, M. (1988). Algorithms and auxiliary routines for FLOW3D release 2.,
Ph.D. Thesis, City University, London, TRERC Res. Memo No. TF/03/88/01.
Ciofalo, M. (1996). Large eddy simulations of turbulent flow with heat transfer in
simple and complex geometries using Harwell-Flow3D., Applied Mathematical
Modelling, Vol. 20, pp. 262-271.
Ciofalo, M. and Collins. M. W. (1989). Large eddy simulations of turbulent flow in
plane and rib-roughened channels., Proceedings of the 6
th
International Conference:
Numerical Methods in Laminar and Turbulent Flows, Swansea, UK, July 1989.
Clarke, D. S. and Wilkes, N. S. (1989). The calculation of turbulent flows in
complex geometries using the differential stress model., AERE - R 13428, Technical
Report, CFDS, Harwell, Oxford, England.
Coakley, T. J. (1993). Turbulence modelling for the compressible Navier-Stokes
Equations., AIAA paper, Vol. 83, p. 1693.
Cook, N. J. (1986a). The designers guide to wind loading on building structures:
part 1., Butterworths , London, England.
137
Cook, N. J. (1986b). The designers guide to wind loading on building structures part
2., Butterworths, London, England.
CP3 Chapter V Part II. (1972). Code of basic data for the design of buildings.,
HMSO, London, England.
Craft, T. J. and Launder, B. E. (1992). New wall reflection model applied to the
turbulent impinging jet., AIAA Journal, Vol. 30, p. 2970.
Craft, T. J., Graham, A. and Launder, B. E (1993b). Impinging jet studies for
turbulence model assessments Part 2: An examination of the performance of 4
turbulence models., International. Journal of Heat and Mass Transfer, Vol. 36 p.
2685.
Craft, T. J, Launder, and Suga, K. (1993a). Extending the applicability of the eddy
viscosity model through the use of deformation invarients and non linear elements.,
Proceedings of the 5th International Symposium on Refined flow Modelling and,
Turbulence Measurements, p. 125, Paris, France.
Craft, T. J., Launder, B. E. and Suga, K. (1996). Development and application of a
cubic eddy-viscosity model of turbulence., International Journal of Heat and Fluid
Flow, Vol. 17, pp. 108-115.
Davenport, A.G. (1999). The Missing Links, The 10
th
ICWE, Copenhagen, June
1999.
Deaves, D. M. and Harris, R. I. (1978). A mathematical model of the structure of
strong winds. CIRIA Report 76. London, Construction Industry Research and
Information Association.
Deissler, T. (1989). On the nature of Navier-Stokes turbulence., NASA, TM
109183.
Delaunay, D., Lakehal, D. and Pierrot, D. (1995). Numerical approach for wind load
prediction on buildings and structures., Journal of Wind Engineering and Industrial
Aerodynamics, Vol. 57, pp. 307-321, Elsevier, Amsterdam, Netherlands .
138
Dickens, M. (1999). Private Communication, Computational Dynamics, Shepherds
Bush, London, UK
Driest, E. R. v. (1956). On turbulent flow near a wall., Journal of the Aeronautical
Sciences, pp. 1007-1011.
Durbin, P. A. (1993) A Reynolds stress model for near wall turbulence., Journal of
Fluid Mechanics, Vol. 249 pp. 465-499, 1993.
Durbin, P. A. and Speziale, C. G. (1991). Local anisotropy in strained turbulent flow
at high Reynolds numbers., Journal of Fluids Engineering, Vol. 113, p. 707.
Eaton, K.J. and Newbury, C.W. (1972) Wind Loading Handbook., BRE Report,
HMSO, London, England.
Ferziger, J. (1993) Simulation of complex turbulent flows: Recent advances and
prospects in wind engineering., Journal of Wind Engineering and Industrial
Aerodynamics, Vol. 46 47, pp. 195-212, Elsevier, Amsterdam, Netherlands.
Ferziger, J. (1994). Lecture notes of ICASE/LaRC Short course on Turbulence
Modelling and Prediction, March 1994 in Gatskim, T. B., Hussaini, M. Y. and
Lumley, J. L, (1996). Simulation and Modeling of Turbulent Flows, Oxford
University Press, Oxford, UK.
FLUENT Users Manual Version 4.22 (1992), Fluent Europe Ltd, Sheffield, UK.
Frank, W. and Mauch, H. (1993) Large-eddy simulation of the flow around building
models., Journal of Wind Engineering and Industrial Aerodynamics, Vol. 46-47 pp.
213-218, Elsevier, Amsterdam, Netherlands.
Fu, S., Launder, B. and Leschziner, M. (1993) Modeling strongly swirling
recirculating jet flow with Reynolds Stress transport closures., Proceedings of the 5th
International Symposium on Refined Flow Modeling and Turbulence Measurements.
Gaskell, P. H. and Lau, A. K. C. (1988). Curvature Compensated Convective
Transport: Smart, A new boundness preserving transport algorithm., International
Journal for Numerical Methods in Fluids, Vol. 8, pp. 617-641, John Wiley and Sons
Ltd., Chichester, UK.
139
Gatski, T. B., Hussaini, M. Y. and Lumley, J. L. (1996). Simulation and modelling of
turbulent flows., Oxford University Press, Oxford, UK.
Germano, M., Piomelli, P., Moin, P. and Cabot, W. H. (1991). A dynamic sub-grid
scale eddy viscosity model., Physics of Fluids A3, Vol. 7, pp. 1769-1765.
Gibson, M. M. and Launder, B. E. (1978). Ground effects on pressure fluctuations in
the atmospheric boundary layer., Journal of Fluid Mechanics, Vol. 86 pt. 3, pp. 491-
511, 1978.
Gleick, J. (1988). Chaos: making a new science., Penguin books, New York.
Goldberg, V. (1987). Separated flow predictions using a new turbulence model.,
Rockwell International Science Centre, California, USA.
Gottlieb, D. and Orszag, S. A. (1977). Numerical analysis of spectral methods:
Theory and Application., SIAM, Philadelphia.
Haggkuist, K., Suensson, U. and Taesler, R. (1989). Numerical simulation of
pressure fields around buildings., Building and Environment, Vol. 24 No. 1 pp. 65-
72.
Hanjalic, K. (1994). Advanced turbulence closure models., International Journal of
Heat and Fluid flow, Vol. 15, p. 178.
Hanjalic, K., Obi, S. and Hadzic, I. (1998). Wall-jets and flows over wall mounted
cubical obstacles., 6
th
ERCROFTAC/IAHR/COST Workshop on Refined Flow
Modelling, ERCROFTAC Bulletin, No. 36, pp. 38-49, March 1998.
Hargreaves, D. M. (1998). Private Communication, The University Of Nottingham,
Nottingham, UK.
Harris, R.I. and Deaves, D.M. (1981). The structure of strong winds. Wind
engineering in the eighties., Proceedings of the CIRIA Conference, London,
Construction Industry Research and Information Association, Paper 4.
Hinze, J.O. (1975). Turbulence., Mcgraw Hill.
140
Hirsch, C. (1990). Numerical Computation of Internal and External Flows, Vol. 1,
p. 493, John Wiley and Sons, Chichester, UK.
Holscher, N. and Niemann, H. J. (1998). Towards quality assurance for wind tunnel
tests: A comparative testing program of the Windtechnologische Gesellschaft.,
Journal of Wind Engineering and Industrial Aerodynamics, Vol. 74-76, pp. 599-608,
Elsevier, Amsterdam, Netherlands.
Hoxey, R.P (1997, 1998, 1999). Private communications, The Silsoe Research
Institute, Silsoe, Bedfordshire, UK.
Hoxey, R.P. and Richards, P.J. (1992) Spectral characteristics of the atmospheric
boundary layer near the ground., Paper presented to the 1
st
UK Wind Engineering
Society Conference, Cambridge, UK.
Hoxey, R.P., Short, L. and Richards, P.J. (1999). Quasi-steady theory developed with
experimental verification., Proceedings of the 10
th
ICWE, Copenhagen 1999, Vol. 3,
pp. 1679-1692.
Hoxey, R.P., Richards, R. P., Richardson, G. M., Robertson, A. P. and Short, J. L.
(1995) The Silsoe Structures Building - The completed experiment part 2., 9th
International Conference on Wind Engineering 1995, New Delhi, India.
Huxye, P. G. and Leschziner, M. A. (1985). Stabilisation of recirculating flow
computations performed with 2nd moment closure and 3rd order discretization.,
Proceedings of the 5th Symposium on Turbulent Shear Flow, pp. 20.7-21.2, Cornell
University.
Iacovides, H. and Launder, B. E. (1987). `The numerical simulation of flow and heat
transfer in tubes in orthagonal mode rotation., Proceedings of the 6th International
Symposium on Turbulent Shear Flows, 1987, Toulouse, France.
Jacobsen, C. B. (1997) Large Eddy Simulation of confined swirling flow: mumerical
part., Ph.D. Thesis submitted to Aalborg University, Institute of Energy Technology.
Jensen, M. (1958). The model law for phenomena in natural wind., Ingenioren
(International Edition), Vol. 2, No. 4.
141
Jones, I. P. (1999). Private Communication, CFX - International, Harwell,
Oxfordshire, UK.
Kawai, H. (1990). A discrete vortex analysis of flow around a vibrating cylinder with
a splitter plate., Journal of Wind Engineering and Industrial Aerodynamics, Vol. 35,
pp. 259-273.
Kawamoto, S. (1996) Improved turbulence models for estimation of wind loading.,
Report for the Nippon Sheet Glass Co., Chiba, Japan.
Kawamoto, S. Kawabata, S. and Tanahashi, T. (1993) Numerical analysis of wind
around buildings using a high speed GSMAC FEM., Journal of Wind Engineering
and Industrial Aerodynamics, Vol. 46-47, pp. 115-120.
Kim, J., Kline, S. J. and Johnston, J. P. (1980). Investigation of a reattaching
turbulent flow over a backward facing step., Transactions of the ASME, Vol. 102,
September 1980.
Kolmogorov, A. N. (1941). de lAcademic des Sci. de 1URSS 30:301.
Lakehal, D. and Rodi, W. (1997). Calculation of flow past a surface mounted cube
with two layer turbulence models., Journal of Wind Engineering and Industrial
Aerodynamics, Vol. 67-68, pp. 65-78.
Lau, A. K. C. (1987). Mathematical modeling of non-reacting and reacting
recirculatory flows., Ph.D. Thesis - Dept. of Mechanical Engineering, The University
of Leeds, Leeds, UK.
Launder, B. E. (1972). Lectures in mathematical models of turbulence., Academic
Press , London, UK.
Launder, B. E. (1994). Lecture notes of ICASE/LaRC Short course on Turbulence
Modelling and Prediction, March 1994 in Gatski, T. B., Hussaini, M. Y. and Lumley,
J. L. (1996). Simulation and modelling of turbulent flows., Oxford University Press,
Oxford, UK.
142
Launder, B. E. and Leschziner, M. A. (1987). Modelling strongly swirling
recirculatory jet flow with Reynolds stress transport closure., Proceedings of the 6th
Symposium on Turbulent Shear Flow, pp. 17.61- 17.6.6, Toulouse, France.
Launder, B.E. and Li, S.P. (1993). On the elimination of wall topography parameters
from second moment closure., Physics of Fluids, Vol. 6 pt.2, pp. 999-1006, 1994,
USA.
Launder, B. E., Reece, G. J. and Rodi, W. (1975) Progress in the development of a
Reynolds Stress turbulence model., Journal of Fluid Mechanics, Vol.68, pt.3, pp.
537-566.
Launder, B. E. and Spalding, D. B. (1974). The numerical computation of turbulent
flows., Computational Methods in Applied Mechanics and Engineering, Vol. 3, pp.
269-289.
Launder, B. E. and Tselepidakis, D. P. (1991). `Directions in second moment
modelling of near wall turbulence.`, AIAA Paper, Vol. 91, p. 219, Reno, Nevada,
USA.
Le, H. Moin, P. and Kim, J. (1993). Direct numerical simulation of turbulent flow
over a backward facing step., Ninth Symposium on Turbulent Shear Flows.
Lee, M. J., Kim, J. and Moin, P. (1990). Structure of turbulence at high shear rates.,
Journal of Fluid Mechanics, Vol. 216, pp. 561-583.
Leonard, A. (1974). Energy cascade in large eddy simulations of turbulent fluid
flows., Advances in Geophysics, Vol. 18A, pp. 237-248.
Leonard, A. (1985). Computation of 3D incompressible flows with vortex elements.,
Annual Review of Fluid Mechanics, Vol.17, pp. 523-529.
Leonard, B. P. (1979). A stable and accurate convective modelling procedure based
on quadratic upstream interpolation., Computational Methods in Applied Mechanics
and Engineering, Vol. 23, pp. 293-312.
143
Leschziner, M. A. (1990) Modelling engineering flows with Reynolds stress
turbulence closure., Journal of Wind Engineering and Industrial Aerodynamics, Vol.
35, pp. 21-47, Elsevier, Amsterdam, Netherlands.
Leschziner, M.A. (1993). Computational modelling of complex turbulent flow,
expectations, reality and prospects., Journal of Wind Engineering and Industrial
Aerodynamics, Vol. 46-47, pp. 37-51, Elsevier, Amsterdam, Netherlands.
Leschziner, M. A. (1995). Modelling turbulence in physically complex flows.,
Industrial Hydraulics and Multiphase flows - Hydro 2000, Thomas Telford, London,
UK.
Libby, P. A. (1996). An introduction to turbulence., Taylor and Francis, London,
UK.
Livesey, F. M. and Mikkelsen, A. C. (1995). Evaluation of the use of the numerical
k- model Kameleon II for predicting wind pressures on building surfaces., Journal
of Wind Engineering and Industrial Aerodynamics, Vol. 57, pp. 375-389, Elsevier,
Amsterdam, Netherlands.
Ludwig, G. R. and Sundaram, T. R. (1969) On the laboratory simulation of small
scale atmospheric turbulence., CAL report, 1969.
Lumley, J. (1970). Towards a turbulent constitutive relation., Journal of Fluid
Mechanics, Vol. 41, p. 413.
Lumley, J., Zhu, J. and Shih, T. (1995). A new Reynolds Stress algebraic equation
model., Computational Methods in Applied Mechanics and Engineering, Vol.125 pp.
287-302.
Mac Comb, W. D. (1990). The physics of fluid turbulence., Oxford Science
Publishers, Clarendon Press, Oxford, UK.
Malalasekera, W. and Versteeg, H. K. (1995). An introduction to computational fluid
mechanics., Longman Scientific and Technical, Essex, UK.
144
Malecki, P. Cousteix, J. and Houdeville, R. (1993). 3D boundary layer calculations
with a two layer turbulence model., Proceedings of the 5th International Symposium
on Refined Flow Modeling and Turbulence Measurements.
Mathews, E. H. (1987). Prediction of the wind generated pressure distribution
around buildings., Journal of Wind Engineering and Industrial Aerodynamics, Vol.
25, pp. 219-228.
Melbourne, W. H. (1995). Turbulence and the leading edge phenomena., Dept. of
Mechanical Engineering, Monash University, Clayton, Victoria, Australia.
Menter, R. F. (1992). Influence of freestream values on k- turbulence model
predictions., AIAA Journal, Vol. 30, No. 6, pp. 1657-1659.
Menter, R. F. and Grotjans, H. (1999). Application of advanced turbulence models to
complex industrial flows., AEA Technology GmbH, Straudenfeldweg, 12 D-83624
Otterfing, Germany.
Murakami, S. (1990). Computational wind engineering., Journal of Wind
Engineering and Industrial Aerodynamics, Vol 35, pp. 4-6, Elsevier, Amsterdam,
Netherlands.
Murakami, S. (1993). Comparison of various turbulence models applied to a bluff
body., Journal of Wind Engineering and Industrial Aerodynamics, Vol. 46-47, pp.
21-36, Elsevier, Amsterdam, Netherlands.
Murakami, S. (1997). Overview of turbulence models applied in CWE., Institute of
Industrial Science, University of Tokyo, Japan.
Murakami, S., Kato, S. and Kondo, Y. (1990) Examining the k- EVM by means of
ASM for a 3D horizontal buoyant jet in enclosed space., Engineering Turbulence
Modelling and Experiments, pp. 205-214, Elsevier Science Publishing.
Murakami, S. and Mochida, A. (1988). 3D numerical simulation of air flow around a
cubic model by means of the k- model., Journal of Wind Engineering and Industrial
Aerodynamics, Vol. 31, pp. 283-303, Elsevier, Amsterdam, Netherlands.
145
Murakami, S. and Mochida, A. (1989a). Applications of CFD to bluff body
aerodynamics - state of the art volume., Journal of Wind Engineering and Industrial
Aerodynamics 1989, Elsevier, Amsterdam, Netherlands.
Murakami, S. and Mochida, A. (1989b). 3D numerical simulation of turbulent flow
around buildings using the k- turbulence model., Building and Environment, Vol.
24, No.1, pp. 51-64.
Murakami, S. and Mochida, A. (1999) Past, present and future of CWE: The view
from 1999., The 10
th
ICWE, Copenhagen, Denmark, 1999.
Murakami, S., Mochida, A. and Hayashi, Y. (1990). Examining the k- model by
means of a wind tunnel test and LES of the turbulence structure around a cube.,
Journal of Wind Engineering and Industrial Aerodynamics, Vol. 35, pp. 87-100,
Elsevier, Amsterdam, Netherlands.
Murakami, S. Mochida, A. and Hayasi, Y. (1991) Scrutinizing the k- EVM and ASM
by means of LES and wind tunnel for flow fields around a cube., 8th symposium on
Turbulent Shear Flows, 1991, pp. 17.1-17.2, Technical University of Munich,
Germany, Elsevier Science Publishers.
Murakami, S., Mochida, A. and Hibi, K. (1987). 3D numerical simulation of air flow
around a cubic model by means of LES., Journal of Wind Engineering and Industrial
Aerodynamics, Vol. 25, pp. 291-305, Elsevier, Amsterdam, Netherlands.
Murakami, S., Mochida, A., Hayashi, Y. and Sakamoto, S. (1992) Numerical study
on velocity pressure field and wind forces for bluff bodies by k-, ASM, and LES.,
Journal of Wind Engineering and Industrial Aerodynamics, Vol. 41-44 pp. 2841-
2852, Elsevier, Amsterdam, Netherlands.
Murakami, S., Mochida, A. and Ooka, R. (1993) Numerical simulation of flow fields
over a surface-mounted cube with various second-moment closure models. Ninth
Symposium on Turbulent Shear Flows, Kyoto, Japan, 1993.
Murakami, S., Mochida, A., Ooka, R., Shinsuke, K. and Iizuka, S. (1996). Numerical
prediction of flow around a building with various turbulence models: Comparison of
k- EVM, ASM, DSM and LES with wind tunnel tests., Transactions of the ASHRAE
Symposia, 1996, Vol. 1, pp. 741-753.
146
Nakamura, Y. and Ozone, S. (1987) The effects of turbulence on separated and
reattaching flow., Journal of Fluid Mechanics, Vol. 178, pp. 477-490.
Naot, D., Shavit, A. and Wolfstein, M. (1970) Interactions between components of
the turbulent velocity correlation tensor., Israel Journal of Technology, Vol. 8, p.
259.
Ng, K. H. (1971).Predictions of turbulent boundary-layer developments using a two-
equation model of turbulence., Ph.D Thesis, University of London, UK.
Orzag, S. A. (1994). Lecture notes of ICASE/LaRC Short course on Turbulence
Modelling and Prediction, March 1994, in Gatski, T. B., Hussaini, M.Y. and Lumley,
J.L. (1996). Simulation and modelling of turbulent flows., Oxford University Press,
Oxford, UK.
Parameswaran, S., Kiris, I., Sun, R. and Gleason, M. (1993) Flow structure around a
3D bluff body in ground proximity: a computational study., Journal of Wind
Engineering and Industrial Aerodynamics, Vol. 46-47, pp. 791-800, Elsevier,
Amsterdam, Netherlands .
Patankar, S. V. (1980) Numerical heat transfer and fluid flow., Hemisphere -
McGraw Hill, New York, USA.
Patankar, S. V. and Spalding, D. B. (1972). A calculation procedure for heat, mass
and momentum transfer in three dimensional parabolic flows., International Journal
of Heat Mass Transfer, Vol. 15, p. 1787.
Patel, V. C., Rodi, W. and Scheuerer, G. (1985). `Turbulence models for near wall
and low Reynolds number flows: A review., AIAA Journal, Vol. 23 pp. 1308-1319.
Patel, V. C. and Satiropouls, F. (1995) Turbulence anisotropy and near wall
modelling in predicting 3D shear flows., AIAA Journal, Vol. 33, No.3.
Paterson, D. A. and Apelt, C. J. (1989). Simulation of wind flow around 3D
buildings. Building and Environment, Vol. 24, No.1 pp. 39-50.
147
Perot, B. and Moin, P. (1995). Shear free turbulent boundary layers Part 1, Physical
insights into near wall turbulence., Journal of Fluid Mechanics, Vol. 295 pp. 199-
227.
Piomelli, U. (1994). Large eddy simulation of turbulent flows., Technical Report
767, TAM Report.
Pollard, A. and Martin, R. (1993). A comparative study of 11 models of turbulence.,
Proceedings of the 5th International Symposium on Refined Flow Modelling and
Turbulence Measurements.
Pope, S. B. (1975). A more general effective-viscosity hypothesis., Journal of Fluid
Mechanics, Vol. 72, pt. 2, pp.331-340.
Qasim, A., Parameswaran, S. and Maxwell, T. (1992) Computational predictions of
flow over a 2D building., Journal of Wind Engineering and Industrial Aerodynamics,
Vol. 41-44, pp. 2839-2840, Elsevier, Amsterdam, Netherlands.
Quinn, A. (1997, 1999). Private Communications, The Silsoe Research Institute,
Silsoe, Bedfordshire, UK.
Reynolds, A. J. (1974). Turbulent flow in engineering., J. Wiley and Sons, London,
UK.
Reynolds, W. C. (1987). Fundamentals of turbulence for turbulence modelling and
simulation. , Lecture notes for the Von Karman Institute AGARD, Lecture Series No.
86, pp. 1-66, New York, USA.
Rhie, C. M. and Chow, W. L. (1983). Numerical study of turbulent flow past an
aerofoil with trailing edge separation., AIAA Journal, Vol. 21, No. 11, pp. 1525-
1532.
Richards, P. J. (1989) Computational modelling of wind flows around low rise
buildings using Phoenix., Report for the ARFC Institute of Engineering Research
Wrest Park, Silsoe Research Institute, Bedfordshire, UK.
Richards, P. J. and Hoxey, R. P. (1992) Computational and wind tunnel modelling of
mean wind loads on the Silsoe Structures building., Journal of Wind Engineering and
148
Industrial Aerodynamics, Vol. 41-44, pp. 1641-1652, Elsevier Publishers,
Amsterdam, Netherlands .
Richards, P. J. and Hoxey, R. P. (1993) Appropriate boundary conditions for
computational wind engineering models using the k- turbulence model., Journal of
Wind Engineering and Industrial Aerodynamics, Vol. 46-47, pp. 145-153, Elsevier
Publishers, Amsterdam, Netherlands .
Richards, P. J. and Hoxey, R. P. (1999). Spectral models of the atmospheric surface
layer., Paper submitted to the 10
th
ICWE, Copenhagen, June 1999.
Richardson, G. M, Hoxey, R. P, Robertson, A. P. and Short, J. L. (1995) The Silsoe
Structures Building - The completed experiment part 1., 9th International conference
on Wind Engineering 1995, New Delhi, India.
Richardson, G. M. and Surrey, D. (1992). The Silsoe Building: a comparison of
pressure coefficients and spectra at model and full-scale., Journal of Wind
Engineering and Industrial Aerodynamics, Vol. 44, pp. 1055-1064.
Richards, P. J. and Wanigaratne, B. S. (1993). A comparison of computer and wind
tunnel modelling of turbulence around the Silsoe Structures Building., Journal of
Wind Engineering and Industrial Aerodynamics, Vol. 46-47, pp. 439-447, Elsevier,
Amsterdam, Netherlands.
Robertson, A. P. (1999). Private Communication, The Silsoe Research Institute,
Silsoe, Bedfordshire, UK.
Robertson, A. P. and Glass, A. G. (1988). The Silsoe Structures Building its design,
instrumentation and research facilities., Divisonal Note DN 1482, AFRC Institute of
Engineering Research, Silsoe, UK, p. 1-59.
Robertson, A. P., Hoxey, R. P., Short, J. L., Ferguson, W. A., (1997). Full-scale
measurements and computational predictions of wind loads on free standing walls.,
Journal of Wind Engineering and Industrial Aerodynamics, Vol. 67-68, pp. 639-646.
Robertson, A. P., Hoxey, R. P., Short, J. L., Ferguson, W. A. and Osmond, S. (1996)
Full-scale testing to determine the wind loads on free standing walls., Journal of
Wind Engineering and Industrial Aerodynamics, Vol. 60, pp. 123-137.
149
Rodi, W. (1972) The prediction of free turbulent boundary layers by use of a two
equation model of turbulence., Ph.D. Thesis, University of London, UK.
Rodi, W. (1990) Experience with two-layer models combining the k- model with a
one equation model near the wall., American Institute of Aeronautics and
Astronautics paper, Vol. 91, p. 216.
Rodi, W. (1993) On the simulation of turbulent flow past bluff bodies., Journal of
Wind Engineering and Industrial Aerodynamics, Vol. 46-47, pp. 3-19, Elsevier,
Amsterdam, Netherlands .
Rotta, J. C. (1972) Turbulente Stromungen., B.G. Teubner, Stuttgart.
Scotti, A. and Meneaveau, C. (1993). Generalised Smagorinsky model for
anisotropic grids., Physics of Fluids, Vol. 9, pp. 2306-2308.
Shah, K.B. and Ferziger, J. H. (1997). A fluid mechanicians view of wind
engineering. LES of flow past a cubic obstacle., Journal of Wind Engineering and
Industrial Aerodynamics Vol. 67 and 68, pp. 211-227.
Shaw, C. T. (1989). Using computational fluid dynamics., Prentice Hall, New York,
USA.
Shih, T. H. and Lumley, J. L. (1985) Modeling of pressure correlation terms in the
Reynolds stress and scalar flux equations., Cornell University Technical Report
FDA.
Shih, T. H. and Lumley, J. L. (1993) A critical comparison of second order closure
with direct numerical simulation of homogeneous turbulent shear flow., AIAA
Journal Vol.31, No.4, April 1993.
Sinai, Y. L. (1995) A simple model of upstream boundary conditions in CFD
simulations within the ABL., AEA Technology. Report, Oxfordshire, UK.
Smagorinsky, J. S. (1963). General circulation experiments with primitive equations,
Part 1; Basic experiments, Monthly Weather Review, Vol. 91, pp. 99-164.
150
Smith, G. D. (1978) Numerical solution of partial differential equations - The finite
difference method., Clarendon Press, Oxford, UK.
So, R.M.C., Lai, Y.G. and Zhang, H.S. (1994) Second order near wall turbulence
closure : A Review., Arizona State University, Arizona, USA.
Spalding, D. B. (1972). A novel finite difference formula for differential equations
involving both 1st and second order derivatives., International Journal of Numerical
Methods in Engineering, Vol. 4 p. 551.
Speziale, C. G. (1987). On nonlinear k- and k- models of turbulence., Journal of
Fluid Mechanics, Vol.178, pp. 459-475.
Speziale, C. G. (1991) Analytical methods for the derivation of Reynolds stress
closures in turbulence., Annual Review of Fluid Mechanics, Vol. 23, pp.107-157.
Speziale, C. G. (1994). Lecture notes of ICASE/LaRC Short course on Turbulence
Modeling and Prediction, March 1994, in Gatski, T. B., Hussaini, M. Y. and Lumley,
J. L. (1996). Simulation and modelling of turbulent flows., Oxford University Press,
Oxford, UK.
Speziale, C. G. (1998). Turbulence modelling for time dependent RANS and VLES: A
review., AIAA Journal, Vol. 36, No.2, Feb 1998
Speziale, C. G. and Abid, R. (1995). Near wall integration of Reynolds stress
turbulence closure with near wall damping., AIAA Journal, Vol. 33, No.10.
Speziale, C. G., Abid, R. and Anderson, E. C. (1990). A critical evaluation of 2
equation models for the near wall turbulence., AIAA Paper, Vol. 90, p. 1481.
Speziale, C. G., Abid, R. and Durbin, P. A. S. (1994) On realisability of the Reynolds
Stress turbulence closures., Journal of Science and Computing, Vol. 9, pp. 369-403.
Speziale, C. G. and Bernard, P. S. (1992). The energy decay in self preserving
isotropic turbulence revisited., Journal of Fluid Mechanics, Vol. 241, pp. 645- 667.
151
Speziale, C. G. and Ngo, T. (1988). Numerical solution of turbulent flow past a
backward facing step using a nonlinear k- model, International Journal of
Engineering Science, Vol. 26, No.10, pp. 1099-1112.
Speziale, C. G., Sarker, S. and Gatski, T. (1991). Modelling the pressure strain
correlation of turbulence: an invariant dynamical systems approach., Journal of
Fluid Mechanics, Vol. 227, pp. 1245-272.
Stewart, I. (1989). Does God play dice, the mathematics of Chaos., Blackwell,
Oxford, UK.
Stone, H.L. (1968). 'Iterative solution of implicit approximations of multi-dimensional
partial differential equations', SIAM Journal of Numerical Analysis, Vol. 5, part 3,
pp530-558.
Suga, K. (1996). Development and application of a non-linear eddy viscosity model
sensitized to stress and strain invariants. Ph.D. Thesis, Department of Mechanical
Engineering, UMIST TDF/95/11.
Summers, D. M., Hanson, T. and Wilson, C. B. (1986). Validation of a computer
simulation of wind over a building model., Building and Environment, Vol. 21, pp97-
111.
Taylor, G. I. (1938). Proceedings of the Royal Society. A164:476
Taylor, G. I. (1921). Proceedings of the London Mathematical Society.
Tehrani, A (1999), Private communication, Rolls Royce Nuclear Power,
Wolverhampton, UK.
Tennekes, H. and Lumley, J. L. (1972). A first course in turbulence., The MIT
Press, Cambridge, UK.
Teubner, B. G. (1996). Second International Symposium on Computational Wind
Engineering, Program and Abstracts, 1996, Colorado State University, USA.
Thomas, N. H. and Hancock, P. E. (1977). Grid turbulence near a moving wall.,
Journal of Fluid Mechanics, Vol. 82, pp. 481 496.
152
Thomas, T. G. and Williams, J. J. R. (1992) Development of a parallel code to
simulate skewed flow over a bluff body., Department of Engineering, Queen Mary
and Westfield College, London, UK.
Tsuchiya, M., Mochida, A., Ishida, Y., Murakami, S. and Konda, K. (1996).
Development of a new k- model for flow and pressure field around a bluff body.,
Second International Symposium on Computational Wind Engineering 1996.
Uzkan, T. and Reynolds, W. C. (1967). A shear free turbulent boundary layer.,
Journal of Fluid Mechanics, Vol.28, pp. 803-821.
Versteeg, H. K. and Malalasekera, W. (1996). An introduction to computational fluid
dynamics - The finite volume method., Longman Scientific and Technical, London,
UK.
Webster, R. (1998). 'Efficient algebraic multi-grid solvers with elementary restriction
and prolongation', International Journal of Numerical Methods in Fluids, Vol. 28,
part 2, pp.317-336, 1998.
Wilcox, D. C. (1994). Turbulence modelling for CFD., DCW Industries Inc.,
Canada
Wilcox, D.C. and Rubinesin, M.W. (1980). Progress in turbulence modelling for
complex flow fields including effects of compressibility., NASA TP - 15.7.
Wilkes, N. G. (1999). Private Communication, CFX-International, Harwell,
Oxfordshire, UK.
Windstorm Loss Dimensions of a Natural Hazard (1990). Munich Reinsurance
Company, Munchener Ruckversicherungs-Gesellschaft, Koniginstrafe 107 D-8000
Munchen 40 Federal Republic of Germany.
Wright, N. G. (1997, 1998, 1999). Private Communications, The University of
Nottingham, Nottingham, UK.
153
Wright, N.G., Easom, G.J. and Hoxey, R. (1997). Improved computational models
for wind engineering., Report for the School of Civil Engineering, The University of
Nottingham, England, UK.
Wright, N.G., Easom, G.J. and Hoxey, R. (1998). Improved computational models
for wind engineering., Proceedings of the 4
th
UK Conference on Wind Engineering,
Bristol University, Bristol, England.
Wright, N.G. and Easom, G.J. (1997). Improved computational models for wind
engineering., Proceeding of the 10
th
International Conference on Wind Engineering,
Copenhagen, Denmark.
Wright, N.G. and Easom, G.J. (1999). Comparison of several computational
turbulence models with full-scale measurements of the flow around a building., Paper
accepted for publication in the Journal of Wind and Structures.
Wright, N.G. and Easom, G.J. (1999). Improved computational models for wind
engineering., Paper accepted for publication in the Journal of Wind and Structures.

Wyngaard, J. C. (1976). The atmospheric boundary layer modelling and
measurement., Springer Verlag.
Yakhot, V. and Orszag, S. A. (1986). Renormalisation Group analysis of turbulence:
Basic theory., Journal of Science and Computing, Vol. 1, pp. 3 51.
Yakhot, V., Orsag, S. A., Thamgam, S., Gatski, T. B. and Speziale, C. G. (1992).
Development of turbulence models for shear flows by a double expansion
technique., Physics of Fluids A, Vol. 4, pp. 1510-1520.
Young, A. D. (1989). Boundary layers., BSP Professional Books, Blackwell
Scientific Publications Ltd., Oxford, UK.
Zannetti, P. (1990). Air pollution modelling: Theories, computational methods and
available software., Computational Mechanics Publications, Von Nostrand,
Reinhold, New York, USA.
154
Zhang, Y. Q., Huber, A. H., Arya, S. and Snyder, W. H (1993). Numerical
simulation to determine the effects of incident wind shear and turbulence level on the
flow around a building. , Journal of Wind Engineering and Industrial Aerodynamics,
Vol. 46-47, pp. 129-134, Elsevier, Amsterdam, Netherlands .
Zhang, H. L. Z. and Ko, N. W. M. (1996). Numerical analysis of incompressible flow
over smooth and grooved circular cylinders., Computers and Fluids, Vol. 25, No.3 ,
pp. 263-281.
Zienkiewicz, O.C. and Taylor, R.L. (1989) The finite element method, 4th edition
Vol. 1 basic formulation and linear problems., McGraw Hill, New York, USA.
155
Appendix A Figures and Illustrations.
Figure 1: Details of the atmospheric boundary layer.
Figure 2: The effect of uniform wind profile on a surface mounted cube (Based on
Cook 1986a)
Distribution of
pressure coefficients
Windward face view
Side view Uniform incident wind
pp
Gradient height
Indication of the
velocity profile
with height
Ekman layer
Inertial sub-
layer
Roughness
sub-layer
156
Figure 3: The effect of a boundary layer mean wind profile on a surface mounted cube
(Based on Cook 1986a)
Figure 4: The flow at the side walls (Based on Cook 1986a)
Streamlines
Distribution of pressure
coefficient
Windward face view F.S = front stagnation
point
Incident wind mean
profile
157
Figure 5: The flow over the roof of a surface mounted cube with uniform incident
wind. (Based on Cook 1986a)
Figure 6: The flow over the roof of a cube with ABL incident wind (Based on Cook
1986a)
Separation and recirculation zone
S.P = separation point
(Side View)
Uniform incident wind
Distribution of pressure
coefficient over roof
Plan view
R.P = reattachment point
S.P = separation point
(Side view)
ABL profile
Distribution of pressure
coefficients over roof
Plan view
158
Figure 7: Cube skewed at 45 degrees to the ABL incident wind. (Based on Cook
1986a)
Figure 8: Flow in the wake (side view, Based on Murakami 1990)
Flow structure
Delta wing vortex
Wedge of attached flow
Pressure coef. dist.
Boundary layer incident wind
Recirculation zone
Free shear layer
159
Convention:
Normal stress = ' ' u u Shear stress = ' ' w v
(Normal stress on plane vw (Plane wu, force in positive w direction)
in direction of u)
Figure 9: Reynolds Stresses: convention.
u
U = f(x)
x
x-h x x+h
Figure 10: Finite differences. (Shaw 1989)
160
WW W P E

ww

w

p

E
w e
Figure 11: Control volume notation.
Figure 12: Overshoots using the QUICK method
Figure 13: A comparison of CCCT and QUICK interpolation profiles.
(Gaskell and Lau 1988)
161
Figure 14: The 6m cube (site photograph).
Figure 15: The Silsoe Structures Building (site photograph).
162
Figure 16: The Silsoe Experimental Wall (site photograph).

Figure 17: Positioning of the outlet boundary (Based on Versteeg and Malalasekera
1996)
163
Figure 18: Comparison of boundary layer profiles for the DSM and standard k-
models. (Inlet velocity = 10m/s at 60m).
Figure 19: Comparison of turbulent kinetic energy, boundary layer profiles for the
DSM and standard k- models. (Inlet velocity = 0.1m/s at 60m).
0
10
20
30
40
50
60
70
80
90
0 2 4 6 8 10 12
Velocity (m/s).
H
e
i
g
h
t

a
b
o
v
e

g
r
o
u
n
d

(
m
)
.
Std. k-eps.
D.S.M.
0
10
20
30
40
50
60
70
80
90
100
0 0.00002 0.00004 0.00006 0.00008 0.0001
Turbulent kinetic energy (m
2
/s
2
)
H
e
i
g
h
t

a
b
o
v
e

g
r
o
u
n
d

(
m
)
.
D.S.M.
Std. k-eps.
164
Figure 20: 6m cube, normal orientation, comparison of windward face pressure
distributions for different boundary layer roughness heights and building roughness.
Figure 21: 6m cube, normal orientation, comparison of roof pressure distributions for
different boundary layer roughness heights and building roughness.
-1.2
-1
-0.8
-0.6
-0.4
-0.2
0
0 2 4 6
Distance along roof (m)
P
r
e
s
s
u
r
e

C
o
e
f
f
i
c
i
e
n
t

(
C
p
)
.
RHL=0.04
RHL=0.01
Building RHL=0.001
0
1
2
3
4
5
6
0 0.2 0.4 0.6 0.8 1
Pressure Coefficient. (Cp).
H
e
i
g
h
t

(
m
)
.
RHL=0.04
Build RHL=0.001
RHL=0.01
165
Figure 22: 6m cube, normal orientation, comparison of windward face pressure
distributions for the CFX original and revised wall functions.
Figure 23: 6m cube, normal orientation, comparison of roof pressure distributions for
the CFX original and revised wall functions.
-1.2
-1
-0.8
-0.6
-0.4
-0.2
0
0 1 2 3 4 5 6
Distance along roof (m)
P
r
e
s
s
u
r
e

C
o
e
f
f
i
c
i
e
n
t
.

(
C
p
)
CFX original
Revised
0
1
2
3
4
5
6
0 0.2 0.4 0.6 0.8 1
Pressure Coefficient. (Cp).
H
e
i
g
h
t

(
m
)
.
Revised
CFX original
166
Figure 24: 6m cube, normal orientation, comparisons of windward face pressure
distributions for the grid refinement test number 1.
Figure 25: 6m cube, normal orientation, comparisons of roof pressure distributions for
the grid refinement test number 1.
-1.2
-1
-0.8
-0.6
-0.4
-0.2
0
0 1 2 3 4 5 6
Distance along roof (m).
P
r
e
s
s
u
r
e

C
o
e
f
f
i
c
i
e
n
t
.

(
C
p
)
120,000
550,000
0
1
2
3
4
5
6
0 0.2 0.4 0.6 0.8 1
Pressure Coefficient. (Cp)
H
e
i
g
h
t

(
m
)
.
120,000
550,000
167
Figure 26: 6m cube, normal orientation, comparison of roof pressure distributions for
the grid refinement test number 2 (Note: the key values refer to number of nodes
around cube).
Figure 27: 6m cube, normal orientation, comparison of roof pressure distributions for
the grid refinement test number 2 (Note: the key values refer to number of nodes
around cube).
-1.2
-1
-0.8
-0.6
-0.4
-0.2
0
0 1 2 3 4 5 6
Distance along roof (m)
P
r
e
s
s
u
r
e

c
o
e
f
f
i
c
i
e
n
t
.

(
C
p
)
Experimental
20*20
40*40
-1.2
-1
-0.8
-0.6
-0.4
-0.2
0
0 1 2 3 4 5 6
Distance along roof (m)
P
r
e
s
s
u
r
e

c
o
e
f
f
i
c
i
e
n
t
.

(
C
p
)
Experimental
60*60
40*40
168
Figure 28: 6m cube, skewed orientation, comparison of roof pressure distributions for
the grid refinement tests (Note: the key values refer to number of nodes around cube).
Figure 29: 6m cube, skewed orientation, comparison of roof pressure distributions for
the grid refinement tests (Note: the key values refer to number of nodes around cube).
-1.2
-1
-0.8
-0.6
-0.4
-0.2
0
0 1 2 3 4 5 6
Distance along roof (m).
P
r
e
s
s
u
r
e

c
o
e
f
f
i
c
i
e
n
t
.

(
C
p
)
60*60
90*90
-1.2
-1
-0.8
-0.6
-0.4
-0.2
0
0 1 2 3 4 5 6
Distance along roof (m)
P
r
e
s
s
u
r
e

c
o
e
f
f
i
c
i
e
n
t
.

(
C
p
)
40*40
20*20
169
Figure 30: 6m cube, skewed orientation, comparison of the roof lateral velocity
component, at the centre of the cube (y coordinate plane), for the grid refinement tests
(Note: the key values refer to number of nodes around the cube).
Figure 31: SSB, sharp eaves, windward roof pressure distribution, grid refinement test
(Note: the key values refer to the total number of nodes in domain).
-0.6
-0.4
-0.2
0
0.2
0.4
0.6
0.8
44 46 48
Distance along roof (m)
(Cube extends from 44 to 52m)
U
/
u
r
e
f
.
20*20
90*90
-2
-1.8
-1.6
-1.4
-1.2
-1
-0.8
-0.6
-0.4
-0.2
0
0 1 2 3 4 5
Distance along roof (m).
P
r
e
s
s
u
r
e

C
o
e
f
f
i
c
i
e
n
t
.

(
C
p
)
.
220,000
440,000
750,000
170
Figure 32: Silsoe experimental wall, windward face pressure distribution, grid
refinement test (Note: the key values refer to the total number of nodes in domain).
Figure 33: Convergence history, residual plot to show solution convergence for a
single numerical precision run.
0
0.1
0.2
0.3
0.4
0.5
0.6
0 5 10 15 20
Distance along wall (m).
M
e
a
n

P
r
e
s
s
u
r
e

C
o
e
f
f
i
c
i
e
n
t

(
C
p
)
.
550,000
1million
171
Figure 34: Convergence history, residual plot to show solution convergence for a
double numerical precision run.
Figure 35: 6m cube, normal orientation, roof pressure distribution, Reynolds number
test.
-1.2
-1
-0.8
-0.6
-0.4
-0.2
0
0 1 2 3 4 5 6
Distance along roof (m)
P
r
e
s
s
u
r
e

C
o
e
f
f
i
c
i
e
n
t
.

(
C
p
)
0.1m/s
10m/s
172
Figure 36: 6m cube, normal orientation, roof vortex strength, Reynolds number test.
Figure 37: 6m cube, normal orientation, roof pressure distribution, comparison of
differencing schemes.
-0.045
-0.04
-0.035
-0.03
-0.025
-0.02
-0.015
-0.01
-0.005
0
0 1 2 3 4 5 6
Distance along roof (m).
U
/
u
r
e
f
.
0.1m/s
10m/s
-1.2
-1
-0.8
-0.6
-0.4
-0.2
0
0 1 2 3 4 5 6
Distance along roof (m).
P
r
e
s
s
u
r
e

C
o
e
f
f
i
c
i
e
n
t

(
C
p
)
.
Hybrid
CCCT
173
Figure 38: 6m cube, normal orientation, roof turbulent kinetic energy distribution,
comparison of differencing schemes.
Figure 39: 6m cube, normal orientation, roof vortex strength, comparison of
differencing schemes.
-0.05
-0.045
-0.04
-0.035
-0.03
-0.025
-0.02
-0.015
-0.01
-0.005
0
0 1 2 3 4 5 6
Distance along roof (m).
U
/
u
r
e
f
.
Hybrid
CCCT
0
0.5
1
1.5
2
2.5
3
3.5
4
4.5
5
0 1 2 3 4 5
Distance along roof (m).
T
u
r
b
u
l
e
n
t

k
i
n
e
t
i
c

e
n
e
r
g
y

(
m
2
/
s
2
)
Hybrid
CCCT
174
Figure 40: False diffusion with the hybrid differencing scheme.
Figure 41: False diffusion with the CCCT differencing scheme
175
Figure 42: The computational domain for the 6m cube model.
Figure 43: The CFD domain for the 6m cube simulations.
176
Figure 44: Example of the CFD mesh for the grid refined simulations.
Figure 45: Normal cube measurement location
177
Figure 46: 6m cube, normal orientation, comparison of roof vortex strength and
reattachment point.
Figure 47: 6m cube, normal orientation, comparison of wake vortex strength and
reattachment point.
-0.35
-0.3
-0.25
-0.2
-0.15
-0.1
-0.05
0
0 1 2 3 4 5 6
Distance along roof (m)
U
/
u
r
e
f
Non-linear
MMK
RNG
DSM
-0.035
-0.03
-0.025
-0.02
-0.015
-0.01
-0.005
0
1 2 3 4 5
Normalised wake distance (dist./cube length)
U
/
u
r
e
f
.
Non-linear
MMK
RNG
DSM
Std k-eps
178
Figure 48: Velocity vectors showing the roof flow field for the 6m cube, standard k-
model.
Figure 49: Velocity vectors, the MMK k- model.
179
Figure 50: Velocity vectors, the RNG k- model (arrows shows reattachment location)
Figure 51: Velocity vectors, the non-linear k- model (arrows shows reattachment
location)
180
Figure 52: Velocity vectors, the differential stress model
Figure 53. 6m cube, normal orientation, windward face pressure distribution.
0
1
2
3
4
5
6
0 0.2 0.4 0.6 0.8 1 1.2
Pressure Coefficient. (Cp).
H
e
i
g
h
t

(
m
)
.
Experimental
M.M.K.
Std. k-eps.
RNG
DSM
Non-linear
181
Figure 54: 6m cube, normal orientation, roof pressure distribution.
Figure 55: 6m cube, normal orientation, leeward face pressure distribution.
-2
-1.8
-1.6
-1.4
-1.2
-1
-0.8
-0.6
-0.4
-0.2
0
0 1 2 3 4 5 6
Distance along roof (m).
P
r
e
s
s
u
r
e

C
o
e
f
f
i
c
i
e
n
t
.

(
C
p
)
.
Experimental
Non-linear
DSM
RNG
Std. k-eps
MMK
0
1
2
3
4
5
6
-0.4 -0.3 -0.2 -0.1 0
Pressure Coefficient. (Cp).
H
e
i
g
h
t

(
m
)
.
Experimental
MMK
Std. k-eps
DSM
RNG
Non-linear
182
Figure 56: 6m cube, normal orientation, side face pressure distribution.
Figure 57: 6m cube, normal orientation, lateral roof pressure distribution.
0
1
2
3
4
5
6
-0.8 -0.6 -0.4 -0.2 0
Pressure Coefficient. (Cp).
H
e
i
g
h
t

(
m
)
.
Experimental
MMK
Std. k-eps
DSM
RNG
Non-linear
-0.9
-0.8
-0.7
-0.6
-0.5
-0.4
-0.3
-0.2
-0.1
0
0 1 2 3 4 5 6
Distance across roof (m)
P
r
e
s
s
u
r
e

C
o
e
f
f
i
c
i
e
n
t
.

(
C
p
)
Experiment.
Non-linear
RNG
DSM
Std. k-eps
MMK
183
Figure 58: 6m cube, normal orientation, windward face turbulent kinetic energy
distribution.
Figure 59: 6m cube, normal orientation, roof turbulent kinetic energy distribution
0
1
2
3
4
5
6
7
0 1 2 3 4 5 6
Distance along roof (m).
T
u
r
b
u
l
e
n
t

k
i
n
e
t
i
c

e
n
e
r
g
y

(
m
2
/
s
2
)
.
Std. k-eps.
MMK
Non-linear
DSM
RNG
0
1
2
3
4
5
6
7
0 5 10 15 20
Turbulent kinetic energy (m
2
/s
2
)
H
e
i
g
h
t

(
m
)
.
MMK
Std. k-eps.
RNG
DSM
Non-linear
184
Figure 60: 6m cube, normal orientation, centreline turbulent kinetic energy contours
for the standard k- model.
Figure 61: 6m cube, normal orientation, centreline turbulent kinetic energy contours
for the MMK k- model.
185
Figure 62: 6m cube, normal orientation, centreline turbulent kinetic energy contours
for the RNG k- model.
Figure 63: 6m cube, normal orientation, centreline turbulent kinetic energy contours
for the non-linear k- model.
186
Figure 64: 6m cube, normal orientation, centreline turbulent kinetic energy contours
for the differential stress model.
Figure 65: Skewed cube measurement locations
187
Figure 66: 6m cube, skewed orientation, comparison of roof lateral velocity
component measured across the centre of the cube (y coordinate plane).
Figure 67: Skewed cube velocity vectors showing the delta wing vortices at the centre
of the cube, using the non-linear k- model
-0.5
-0.4
-0.3
-0.2
-0.1
0
0.1
0.2
0.3
0.4
0.5
44.5 45.5 46.5 47.5 48.5
Distance (m)
U
/
u
r
e
f
.
Non-linear
MMK
RNG
DSM
Std. k-eps
188
Figure 68: 6m cube, skewed orientation, comparison of windward face pressure
coefficients (measurement line 1).
Figure 69: 6m cube, skewed orientation, comparison of roof pressure coefficients
(measurement line 2).
0
1
2
3
4
5
6
0 0.1 0.2 0.3 0.4 0.5
Pressure Coefficient. (Cp).
H
e
i
g
h
t

(
m
)
.
Experiment
Std. k-eps.
MMK
RNG
Non-lin
DSM
-1.4
-1.2
-1
-0.8
-0.6
-0.4
-0.2
0
0 1 2 3 4 5
Distance along roof (m).
P
r
e
s
s
u
r
e

C
o
e
f
f
i
c
i
e
n
t

(
C
p
)
.
Non-lin
RNG
DSM
MMK
Std. k-eps.
Experiment
189
Figure 70: 6m cube, skewed orientation, comparison of roof pressure coefficients,
grid refined results.
Figure 71: 6m cube, skewed orientation, comparison of leeward face pressure
coefficients (measurement line 3).
0
1
2
3
4
5
6
-0.6 -0.5 -0.4 -0.3 -0.2 -0.1 0
Pressure Coefficient. (Cp).
H
e
i
g
h
t

(
m
)
.
Experimental
Std. k-eps
MMK
DSM
RNG
Non-lin
-1.4
-1.2
-1
-0.8
-0.6
-0.4
-0.2
0
0 1 2 3 4 5 6
Distance along roof (m).
P
r
e
s
s
u
r
e

c
o
e
f
f
i
c
i
e
n
t
.

(
C
p
)
Experiment
RNG
Non-linear
190
Figure 72: Schematic of the SSB / measurement line locations (all dimensions in m)
Figure 73: SSB, sharp eaves, normal orientation, windward face pressure
(measurement line 3).
0
0.5
1
1.5
2
2.5
3
3.5
4
0 0.2 0.4 0.6
Pressure Coefficient. (Cp).
H
e
i
g
h
t

(
m
)
.
Experimental
Non-linear
DSM
RNG
Std k -eps
191
Figure 74: SSB, sharp eaves, normal orientation, windward face turbulent kinetic
energy (measurement line 3).
Figure 75: SSB, sharp eaves, normal orientation, windward roof pressure
(measurement line 3).
0
0.5
1
1.5
2
2.5
3
3.5
4
0 0.02 0.04 0.06 0.08 0.1
Normalised turbulent kinetic energy (m
2
/s
2
/v
2
)
H
e
i
g
h
t

(
m
)
.
Non-linear
DSM
RNG
Std. k-eps
-2
-1.8
-1.6
-1.4
-1.2
-1
-0.8
-0.6
-0.4
-0.2
0
0 1 2 3 4 5
Distance along roof (m).
P
r
e
s
s
u
r
e

C
o
e
f
f
i
c
i
e
n
t

(
C
p
)
.
Experimental
Std k-eps
RNG
DSM
Non-linear
192
Figure 76: SSB, sharp eaves, normal orientation, windward roof turbulent kinetic
energy (measurement line 3).
Figure 77: SSB, shape eaves, normal orientation, centreline turbulent kinetic energy
contour plot for the non-linear k- model.
0
0.02
0.04
0.06
0.08
0.1
0.12
0.14
0.16
0 1 2 3 4 5 6
Distance along roof. (m).
N
o
r
m
a
l
i
s
e
d

t
u
r
b
u
l
e
n
t

k
i
n
e
t
i
c

e
n
e
r
g
y


(
m
2
/
s
2
/
v
2
)
Std. k-eps
RNG
DSM
Non-linear
193
Figure 78: SSB, sharp eaves, normal orientation, leeward roof pressure distribution
(measurement line 3).
Figure 79: SSB, sharp eaves, normal orientation, leeward face pressure distribution
(measurement line 3).
-1.2
-1
-0.8
-0.6
-0.4
-0.2
0
0 1 2 3 4 5 6
Distance along roof (m).
P
r
e
s
s
u
r
e

C
o
e
f
f
i
c
i
e
n
t
.

(
C
p
)
.
Experimental
Std. k-eps.
RNG
DSM
Non-linear
0
0.5
1
1.5
2
2.5
3
3.5
4
-0.2 -0.15 -0.1 -0.05 0
Pressure Coefficient. (Cp).
H
e
i
g
h
t

(
m
)
.
Experimental
Non-linear
DSM
RNG
Std. k-eps.
194
Figure 80: SSB, curved eaves, normal orientation, windward face pressure
distribution (measurement line 3).
Figure 81: SSB, curved eaves, normal orientation, windward roof pressure
distribution (measurement line 3).
0
0.5
1
1.5
2
2.5
3
3.5
4
0 0.1 0.2 0.3 0.4 0.5
Pressure Coefficient. (Cp).
H
e
i
g
h
t

(
m
)
Experimental
Non-linear
DSM
RNG
Std k-eps.
-1.5
-1.3
-1.1
-0.9
-0.7
-0.5
-0.3
-0.1
0 1 2 3 4
Distance along roof (m).
P
r
e
s
s
u
r
e

C
o
e
f
f
i
c
i
e
n
t
.


(
C
p
)
.
experiment
RNG
Std k-eps.
DSM
Non-linear
195
Figure 82: SSB, curved eaves, normal orientation, leeward roof pressure distribution
(measurement line 3).
Figure 83: SSB, curved eaves, normal orientation, leeward roof turbulent kinetic
energy distribution (measurement line 3).
-1.4
-1.2
-1
-0.8
-0.6
-0.4
-0.2
0
0 1 2 3 4 5 6
Distance along roof (m)
P
r
e
s
s
u
r
e

C
o
e
f
f
I
c
i
e
n
t
.

(
C
p
)
Experimental
Std. k-eps
RNG
DSM
Non-linear
0
0.01
0.02
0.03
0.04
0.05
0.06
0 1 2 3 4 5 6
Distance along roof (m)
N
o
r
m
a
l
i
s
e
d

t
u
r
b
u
l
e
n
t

k
i
n
e
t
i
c

e
n
e
r
g
y

(
m
2
/
s
2
/
v
2
)
Std. k-eps
RNG
DSM
Non-linear
196
Figure 84: SSB, curved eaves, normal orientation, leeward face pressure distribution
(measurement line 3).
Figure 85: SSB, curved eaves, transverse orientation, side face pressure distribution
(measurement line 1).
0
0.5
1
1.5
2
2.5
3
3.5
4
-0.3 -0.25 -0.2 -0.15 -0.1 -0.05 0
Pressure Coefficient. (Cp).
H
e
i
g
h
t

(
m
)
.
Experimental
Non-linear
DSM
RNG
Std. k-eps.
0
0.5
1
1.5
2
2.5
3
3.5
4
-1 -0.8 -0.6 -0.4 -0.2 0
Pressure Coefficient. (Cp).
H
e
i
g
h
t

(
m
)
.
Experimental
Non-linear
RNG
Std. k-eps
197
Figure 86: SSB, curved eaves, transverse orientation, roof pressure distribution
(measurement line 1).
Figure 87: SSB, curved eaves, transverse orientation, centreline velocity
vector plot for the non-linear k- model.
-1.6
-1.4
-1.2
-1
-0.8
-0.6
-0.4
-0.2
0
0 1 2 3 4 5
Distance across roof (from eaves to ridge) (m)
P
r
e
s
s
u
r
e

C
o
e
f
f
i
c
i
e
n
t
.


(
C
p
)
Experiment
Std. k-eps
RNG
Non-linear
198
Figure 88: SSB, curved eaves, transverse orientation, side face pressure distribution
(measurement line 2).
Figure 89: SSB, curved eaves, transverse orientation, roof pressure distribution
(measurement line 2).
0
0.5
1
1.5
2
2.5
3
3.5
4
-1 -0.8 -0.6 -0.4 -0.2 0
Pressure Coefficient. (Cp).
H
e
i
g
h
t

(
m
)
.
Experiment
Non-linear
RNG
Std. k-eps
-0.7
-0.65
-0.6
-0.55
-0.5
-0.45
-0.4
-0.35
-0.3
0 1 2 3 4 5
Distance across roof (from eaves to ridge). (m)
P
r
e
s
s
u
r
e

C
o
e
f
f
i
c
i
e
n
t

(
C
p
)
Experimental
Std. k-eps
RNG
Non-linear
199
Figure 90: SSB, curved eaves, transverse orientation, side face pressure distribution
(measurement line 3).
Figure 91: SSB, curved eaves, transverse orientation, roof pressure distribution
(measurement line 3).
0
0.5
1
1.5
2
2.5
3
3.5
4
-0.2 -0.15 -0.1 -0.05 0
Pressure Coeficient. (Cp).
H
e
i
g
h
t

(
m
)
.
Experimental
Non-linear
RNG
Std. k-eps
-0.25
-0.2
-0.15
-0.1
-0.05
0
0 1 2 3 4 5 6
Distance across roof (from eaves to ridge). (m).
P
r
e
s
s
u
r
e

C
o
e
f
f
i
c
i
e
n
t
.


(
C
p
)
Experimental
Std. k-eps
RNG
Non-linear
200
Figure 92: Schematic of the modular panel layout of the experimental wall (0
incident wind)
Figure 93: Silsoe experimental wall, windward face pressure distribution (0 case).
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0 5 10 15 20
Distance along wall (m).
M
e
a
n

P
r
e
s
s
u
r
e

C
o
e
f
f
i
c
i
e
n
t
.

(
C
p
)
.
Experimental
Non-linear
Std. k-eps.
RNG
201
Figure 94: Silsoe experimental wall, leeward face pressure distribution (0 case).
Figure 95: Silsoe experimental wall, windward face pressure distributions (30 case).
-1
-0.9
-0.8
-0.7
-0.6
-0.5
-0.4
-0.3
-0.2
-0.1
0
0 5 10 15
Distance along wall (m).
M
e
a
n

p
r
e
s
s
u
r
e

c
o
e
f
f
.

(
C
p
)
Non-linear
St. k-eps
RNG
Experimental
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
0 5 10 15
Distance along wall (m)
M
e
a
n

P
r
e
s
s
u
r
e

C
o
e
f
f
i
c
i
e
n
t

(
C
p
)
.
Non-linear
RNG
Std. k-eps.
Experimental
202
Figure 96: Silsoe experimental wall, leeward face pressure distributions (30 case).
Figure 97: Turbulent kinetic energy contours for the experimental wall, the standard
k- model (30 case). (Note: wall may appear to be at an incorrect angle due to page
sizing requirements).
-3
-2.5
-2
-1.5
-1
-0.5
0
0 5 10 15
Distance along wall (m).
M
e
a
n

P
r
e
s
s
u
r
e

C
o
e
f
f
i
c
i
e
n
t

(
C
p
)
.
Non-linear
RNG
Std. k-eps.
Experimental
203
Figure 98: Turbulent kinetic energy contours for the experimental wall, the non-linear
k- model (30 case). (Note: wall may appear to be at an incorrect angle due to page
sizing requirements).
Figure 99: Velocity vectors for the experimental wall using the standard k- model
(30 case). (Note: wall may appear to be at an incorrect angle due to page sizing
requirements).
204
Figure 100: Velocity vectors for the experimental wall using the non-linear k- model
(30 case). (Note: wall may appear to be at an incorrect angle due to page sizing
requirements).
Figure 101: Silsoe experimental wall, windward face pressure distributions (60 case).
0
0.1
0.2
0.3
0.4
0.5
0.6
0 5 10 15
Distance along wall (m).
M
e
a
n

P
r
e
s
s
u
r
e

C
o
e
f
f
i
c
i
e
n
t

(
C
p
)
.
Non-linear
Std. k-eps.
RNG
Experimental
205
Figure 102: Silsoe experimental wall, leeward face pressure distributions (60 case).
Figure 103: Velocity vectors for the experimental wall using the non-linear k- model
(60 case). (Note: wall may appear to be at an incorrect angle due to page sizing
requirements).
-1.6
-1.4
-1.2
-1
-0.8
-0.6
-0.4
-0.2
0
0 5 10 15
Distance along wall (m).
M
e
a
n

P
r
e
s
s
u
r
e

C
o
e
f
f
i
c
i
e
n
t

(
C
p
)
.
Non-linear
Std. k-eps.
RNG
Experimental
206
Figure 104: 6m cube, centreline uw shear stress contour plot for the non-linear k-
model.
Figure 105: 6m cube, centreline uw shear stress contour plot for the differential
stress model..
207
Figure 106: 6m cube, centreline uu stream-wise normal stress plot for the non-linear
k- model.
Figure 107: 6m cube, centreline vv lateral normal stress plot for the non-linear k-
model.
208
Figure 108: 6m cube, centreline ww vertical normal stress plot for the non-linear k-
model.
Figure 109: 6m cube, centreline uu stream-wise normal stress plot for the linear
(standard) k- model.
209
Figure 110: 6m cube, centreline vv stream-wise normal stress plot for the linear
(standard) k- model.
Figure 111: 6m cube, centreline ww stream-wise normal stress plot for the linear
(standard) k- model.
210
Figure 112: Normal Reynolds stress distribution for the roof (centre) of the 6m cube
calculated using the non-linear k- model.
Figure 113: Normal Reynolds stress distribution for the roof (centre) of the 6m cube
calculated using the differential stress model.
0
1
2
3
4
5
6
7
0 1 2 3 4 5 6
Distance along roof (m).
R
e
y
n
o
l
d
s

s
t
r
e
s
s

(
m
2
/
s
2
)
u'u'
w'w'
v'v'
0
1
2
3
4
5
6
0 1 2 3 4 5 6
Distance along roof (m)
R
e
y
n
o
l
d
s

s
t
r
e
s
s

(
m
2
/
s
2
)
w'w'
v'v'
u'u'
211
Figure 114: Normal Reynolds stress distribution for the wake of the 6m cube
calculated using the non-linear k- model
Figure 115: The computational domain for the LES simulations
0
1
2
3
4
5
6
7
8
0 1 2 3 4 5
Normalised wake distance (dist./cube length).
R
e
y
n
o
l
d
s

s
t
r
e
s
s

(
m
2
/
s
2
)
v'v'
w'w'
u'u'
212
Figure 116: LES, time series of fluctuating velocity for the 1/30 scale cube model.
Figure 117: LES transient velocity vector plot number 1, 1/30 scale cube model,
coarse grid (plan view).
-1
-0.5
0
0.5
1
1.5
2
0 200 400 600 800 1000 1200 1400
Time step.
V
e
l
o
c
i
t
y

(
m
/
s
)
.
u (m/s)
213
Figure 118: LES transient plot number 2.
Figure 119: LES transient plot number 3.
214
Figure 120: LES transient plot number 4.
Figure 121: LES, Frequency Spectra for the 1/30 scale cube model.
0.00001
0.0001
0.001
0.01
0.1
1
10
1 10 100
Frequency (Hz).
S
p
e
c
t
r
a
l

D
e
n
s
i
t
y

(
S
n
)
.
215
Figure 122: LES transient flow, velocity vector plot, fine resolution grid, 1/30 scale
cube model (plan view).
Figure 123: LES transient flow, velocity vector plot, 1/30 scale SSB model (side
view).
216
Figure 124: LES transient flow, velocity vector plot, 1/30 scale SSB model, full
domain (side view).
Figure 125: LES transient flow, velocity vector plot, 1/30 scale cube model, coarse
grid, domain and flow field (plan view).
217
Figure 126: LES transient flow, velocity vector plot, coarse grid (side view).

Figure 127: LES, time averaged velocity contour plot, 1/30 scale cube model, coarse
grid (side view).
218

Figure 128: LES, time averaged pressure contour plot, 1/30 scale cube model, coarse
grid (side view).

Figure 129: Roof velocity vectors for the LES comparisons predicted using the
(steady state) non-linear k- model.
219
Figure 130: Roof velocity vectors for the LES comparisons predicted using the
(steady state) low Reynolds number k- model. (Note: The same roof flow field was
found using the low Reynolds number k- model.)
Figure 131: Wake and roof velocity vectors for the LES comparisons predicted using
the (steady state) non-linear k- model.
220